github/kubernetes
4
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

address issue #1488; clean up linewrap and some minor editing issues in the docs/design/* tree

Browse Source 
Signed-off-by: mikebrow <brownwm@us.ibm.com>
This commit is contained in:
mikebrow 2016-04-13 19:55:22 -05:00
parent 4638f2f355
commit 6bdc0bfdb7
39 changed files with 3744 additions and 2375 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

54
docs/design/README.md
View File

@ -34,19 +34,59 @@ Documentation for other releases can be found at
# Kubernetes Design Overview
Kubernetes is a system for managing containerized applications across multiple hosts, providing basic mechanisms for deployment, maintenance, and scaling of applications.
Kubernetes is a system for managing containerized applications across multiple
hosts, providing basic mechanisms for deployment, maintenance, and scaling of
applications.
Kubernetes establishes robust declarative primitives for maintaining the desired state requested by the user. We see these primitives as the main value added by Kubernetes. Self-healing mechanisms, such as auto-restarting, re-scheduling, and replicating containers require active controllers, not just imperative orchestration.
Kubernetes establishes robust declarative primitives for maintaining the desired
state requested by the user. We see these primitives as the main value added by
Kubernetes. Self-healing mechanisms, such as auto-restarting, re-scheduling, and
replicating containers require active controllers, not just imperative
orchestration.
Kubernetes is primarily targeted at applications composed of multiple containers, such as elastic, distributed micro-services. It is also designed to facilitate migration of non-containerized application stacks to Kubernetes. It therefore includes abstractions for grouping containers in both loosely coupled and tightly coupled formations, and provides ways for containers to find and communicate with each other in relatively familiar ways.
Kubernetes is primarily targeted at applications composed of multiple
containers, such as elastic, distributed micro-services. It is also designed to
facilitate migration of non-containerized application stacks to Kubernetes. It
therefore includes abstractions for grouping containers in both loosely coupled
and tightly coupled formations, and provides ways for containers to find and
communicate with each other in relatively familiar ways.
Kubernetes enables users to ask a cluster to run a set of containers. The system automatically chooses hosts to run those containers on. While Kubernetes's scheduler is currently very simple, we expect it to grow in sophistication over time. Scheduling is a policy-rich, topology-aware, workload-specific function that significantly impacts availability, performance, and capacity. The scheduler needs to take into account individual and collective resource requirements, quality of service requirements, hardware/software/policy constraints, affinity and anti-affinity specifications, data locality, inter-workload interference, deadlines, and so on. Workload-specific requirements will be exposed through the API as necessary.
Kubernetes enables users to ask a cluster to run a set of containers. The system
automatically chooses hosts to run those containers on. While Kubernetes's
scheduler is currently very simple, we expect it to grow in sophistication over
time. Scheduling is a policy-rich, topology-aware, workload-specific function
that significantly impacts availability, performance, and capacity. The
scheduler needs to take into account individual and collective resource
requirements, quality of service requirements, hardware/software/policy
constraints, affinity and anti-affinity specifications, data locality,
inter-workload interference, deadlines, and so on. Workload-specific
requirements will be exposed through the API as necessary.
Kubernetes is intended to run on a number of cloud providers, as well as on physical hosts.
Kubernetes is intended to run on a number of cloud providers, as well as on
physical hosts.
A single Kubernetes cluster is not intended to span multiple availability zones. Instead, we recommend building a higher-level layer to replicate complete deployments of highly available applications across multiple zones (see [the multi-cluster doc](../admin/multi-cluster.md) and [cluster federation proposal](../proposals/federation.md) for more details).
A single Kubernetes cluster is not intended to span multiple availability zones.
Instead, we recommend building a higher-level layer to replicate complete
deployments of highly available applications across multiple zones (see
[the multi-cluster doc](../admin/multi-cluster.md) and [cluster federation proposal](../proposals/federation.md)
for more details).
Finally, Kubernetes aspires to be an extensible, pluggable, building-block OSS platform and toolkit. Therefore, architecturally, we want Kubernetes to be built as a collection of pluggable components and layers, with the ability to use alternative schedulers, controllers, storage systems, and distribution mechanisms, and we're evolving its current code in that direction. Furthermore, we want others to be able to extend Kubernetes functionality, such as with higher-level PaaS functionality or multi-cluster layers, without modification of core Kubernetes source. Therefore, its API isn't just (or even necessarily mainly) targeted at end users, but at tool and extension developers. Its APIs are intended to serve as the foundation for an open ecosystem of tools, automation systems, and higher-level API layers. Consequently, there are no "internal" inter-component APIs. All APIs are visible and available, including the APIs used by the scheduler, the node controller, the replication-controller manager, Kubelet's API, etc. There's no glass to break -- in order to handle more complex use cases, one can just access the lower-level APIs in a fully transparent, composable manner.
Finally, Kubernetes aspires to be an extensible, pluggable, building-block OSS
platform and toolkit. Therefore, architecturally, we want Kubernetes to be built
as a collection of pluggable components and layers, with the ability to use
alternative schedulers, controllers, storage systems, and distribution
mechanisms, and we're evolving its current code in that direction. Furthermore,
we want others to be able to extend Kubernetes functionality, such as with
higher-level PaaS functionality or multi-cluster layers, without modification of
core Kubernetes source. Therefore, its API isn't just (or even necessarily
mainly) targeted at end users, but at tool and extension developers. Its APIs
are intended to serve as the foundation for an open ecosystem of tools,
automation systems, and higher-level API layers. Consequently, there are no
"internal" inter-component APIs. All APIs are visible and available, including
the APIs used by the scheduler, the node controller, the replication-controller
manager, Kubelet's API, etc. There's no glass to break -- in order to handle
more complex use cases, one can just access the lower-level APIs in a fully
transparent, composable manner.
For more about the Kubernetes architecture, see [architecture](architecture.md).

289
docs/design/access.md
View File

@ -34,23 +34,30 @@ Documentation for other releases can be found at
# K8s Identity and Access Management Sketch
This document suggests a direction for identity and access management in the Kubernetes system.
This document suggests a direction for identity and access management in the
Kubernetes system.
## Background
High level goals are:
- Have a plan for how identity, authentication, and authorization will fit in to the API.
- Have a plan for partitioning resources within a cluster between independent organizational units.
- Have a plan for how identity, authentication, and authorization will fit in
to the API.
- Have a plan for partitioning resources within a cluster between independent
organizational units.
- Ease integration with existing enterprise and hosted scenarios.
### Actors
Each of these can act as normal users or attackers.
- External Users: People who are accessing applications running on K8s (e.g. a web site served by webserver running in a container on K8s), but who do not have K8s API access.
- K8s Users : People who access the K8s API (e.g. create K8s API objects like Pods)
- External Users: People who are accessing applications running on K8s (e.g.
a web site served by webserver running in a container on K8s), but who do not
have K8s API access.
- K8s Users: People who access the K8s API (e.g. create K8s API objects like
Pods)
- K8s Project Admins: People who manage access for some K8s Users
- K8s Cluster Admins: People who control the machines, networks, or binaries that make up a K8s cluster.
- K8s Cluster Admins: People who control the machines, networks, or binaries
that make up a K8s cluster.
- K8s Admin means K8s Cluster Admins and K8s Project Admins taken together.
### Threats
@ -58,22 +65,31 @@ Each of these can act as normal users or attackers.
Both intentional attacks and accidental use of privilege are concerns.
For both cases it may be useful to think about these categories differently:
- Application Path - attack by sending network messages from the internet to the IP/port of any application running on K8s. May exploit weakness in application or misconfiguration of K8s.
- Application Path - attack by sending network messages from the internet to
the IP/port of any application running on K8s. May exploit weakness in
application or misconfiguration of K8s.
- K8s API Path - attack by sending network messages to any K8s API endpoint.
- Insider Path - attack on K8s system components. Attacker may have privileged access to networks, machines or K8s software and data. Software errors in K8s system components and administrator error are some types of threat in this category.
- Insider Path - attack on K8s system components. Attacker may have
privileged access to networks, machines or K8s software and data. Software
errors in K8s system components and administrator error are some types of threat
in this category.
This document is primarily concerned with K8s API paths, and secondarily with Internal paths. The Application path also needs to be secure, but is not the focus of this document.
This document is primarily concerned with K8s API paths, and secondarily with
Internal paths. The Application path also needs to be secure, but is not the
focus of this document.
### Assets to protect
External User assets:
- Personal information like private messages, or images uploaded by External Users.
- Personal information like private messages, or images uploaded by External
Users.
- web server logs.
K8s User assets:
- External User assets of each K8s User.
- things private to the K8s app, like:
- credentials for accessing other services (docker private repos, storage services, facebook, etc)
- credentials for accessing other services (docker private repos, storage
services, facebook, etc)
- SSL certificates for web servers
- proprietary data and code
@ -82,38 +98,51 @@ K8s Cluster assets:
- Machine Certificates or secrets.
- The value of K8s cluster computing resources (cpu, memory, etc).
This document is primarily about protecting K8s User assets and K8s cluster assets from other K8s Users and K8s Project and Cluster Admins.
This document is primarily about protecting K8s User assets and K8s cluster
assets from other K8s Users and K8s Project and Cluster Admins.
### Usage environments
Cluster in Small organization:
- K8s Admins may be the same people as K8s Users.
- few K8s Admins.
- prefer ease of use to fine-grained access control/precise accounting, etc.
- Product requirement that it be easy for potential K8s Cluster Admin to try out setting up a simple cluster.
- Few K8s Admins.
- Prefer ease of use to fine-grained access control/precise accounting, etc.
- Product requirement that it be easy for potential K8s Cluster Admin to try
out setting up a simple cluster.
Cluster in Large organization:
- K8s Admins typically distinct people from K8s Users. May need to divide K8s Cluster Admin access by roles.
- K8s Admins typically distinct people from K8s Users. May need to divide
K8s Cluster Admin access by roles.
- K8s Users need to be protected from each other.
- Auditing of K8s User and K8s Admin actions important.
- flexible accurate usage accounting and resource controls important.
- Flexible accurate usage accounting and resource controls important.
- Lots of automated access to APIs.
- Need to integrate with existing enterprise directory, authentication, accounting, auditing, and security policy infrastructure.
- Need to integrate with existing enterprise directory, authentication,
accounting, auditing, and security policy infrastructure.
Org-run cluster:
- organization that runs K8s master components is same as the org that runs apps on K8s.
- Organization that runs K8s master components is same as the org that runs
apps on K8s.
- Nodes may be on-premises VMs or physical machines; Cloud VMs; or a mix.
Hosted cluster:
- Offering K8s API as a service, or offering a Paas or Saas built on K8s.
- May already offer web services, and need to integrate with existing customer account concept, and existing authentication, accounting, auditing, and security policy infrastructure.
- May want to leverage K8s User accounts and accounting to manage their User accounts (not a priority to support this use case.)
- Precise and accurate accounting of resources needed. Resource controls needed for hard limits (Users given limited slice of data) and soft limits (Users can grow up to some limit and then be expanded).
- May already offer web services, and need to integrate with existing customer
account concept, and existing authentication, accounting, auditing, and security
policy infrastructure.
- May want to leverage K8s User accounts and accounting to manage their User
accounts (not a priority to support this use case.)
- Precise and accurate accounting of resources needed. Resource controls
needed for hard limits (Users given limited slice of data) and soft limits
(Users can grow up to some limit and then be expanded).
K8s ecosystem services:
- There may be companies that want to offer their existing services (Build, CI, A/B-test, release automation, etc) for use with K8s. There should be some story for this case.
- There may be companies that want to offer their existing services (Build, CI,
A/B-test, release automation, etc) for use with K8s. There should be some story
for this case.
Pods configs should be largely portable between Org-run and hosted configurations.
Pods configs should be largely portable between Org-run and hosted
configurations.
# Design
@ -123,65 +152,99 @@ Related discussion:
- http://issue.k8s.io/443
This doc describes two security profiles:
- Simple profile: like single-user mode. Make it easy to evaluate K8s without lots of configuring accounts and policies. Protects from unauthorized users, but does not partition authorized users.
- Enterprise profile: Provide mechanisms needed for large numbers of users. Defense in depth. Should integrate with existing enterprise security infrastructure.
- Simple profile: like single-user mode. Make it easy to evaluate K8s
without lots of configuring accounts and policies. Protects from unauthorized
users, but does not partition authorized users.
- Enterprise profile: Provide mechanisms needed for large numbers of users.
Defense in depth. Should integrate with existing enterprise security
infrastructure.
K8s distribution should include templates of config, and documentation, for simple and enterprise profiles. System should be flexible enough for knowledgeable users to create intermediate profiles, but K8s developers should only reason about those two Profiles, not a matrix.
K8s distribution should include templates of config, and documentation, for
simple and enterprise profiles. System should be flexible enough for
knowledgeable users to create intermediate profiles, but K8s developers should
only reason about those two Profiles, not a matrix.
Features in this doc are divided into "Initial Feature", and "Improvements". Initial features would be candidates for version 1.00.
Features in this doc are divided into "Initial Feature", and "Improvements".
Initial features would be candidates for version 1.00.
## Identity
### userAccount
K8s will have a `userAccount` API object.
- `userAccount` has a UID which is immutable. This is used to associate users with objects and to record actions in audit logs.
- `userAccount` has a name which is a string and human readable and unique among userAccounts. It is used to refer to users in Policies, to ensure that the Policies are human readable. It can be changed only when there are no Policy objects or other objects which refer to that name. An email address is a suggested format for this field.
- `userAccount` is not related to the unix username of processes in Pods created by that userAccount.
- `userAccount` has a UID which is immutable. This is used to associate users
with objects and to record actions in audit logs.
- `userAccount` has a name which is a string and human readable and unique among
userAccounts. It is used to refer to users in Policies, to ensure that the
Policies are human readable. It can be changed only when there are no Policy
objects or other objects which refer to that name. An email address is a
suggested format for this field.
- `userAccount` is not related to the unix username of processes in Pods created
by that userAccount.
- `userAccount` API objects can have labels.
The system may associate one or more Authentication Methods with a
`userAccount` (but they are not formally part of the userAccount object.)
In a simple deployment, the authentication method for a
user might be an authentication token which is verified by a K8s server. In a
more complex deployment, the authentication might be delegated to
another system which is trusted by the K8s API to authenticate users, but where
the authentication details are unknown to K8s.
In a simple deployment, the authentication method for a user might be an
authentication token which is verified by a K8s server. In a more complex
deployment, the authentication might be delegated to another system which is
trusted by the K8s API to authenticate users, but where the authentication
details are unknown to K8s.
Initial Features:
- there is no superuser `userAccount`
- `userAccount` objects are statically populated in the K8s API store by reading a config file. Only a K8s Cluster Admin can do this.
- `userAccount` can have a default `namespace`. If API call does not specify a `namespace`, the default `namespace` for that caller is assumed.
- `userAccount` is global. A single human with access to multiple namespaces is recommended to only have one userAccount.
- There is no superuser `userAccount`
- `userAccount` objects are statically populated in the K8s API store by reading
a config file. Only a K8s Cluster Admin can do this.
- `userAccount` can have a default `namespace`. If API call does not specify a
`namespace`, the default `namespace` for that caller is assumed.
- `userAccount` is global. A single human with access to multiple namespaces is
recommended to only have one userAccount.
Improvements:
- Make `userAccount` part of a separate API group from core K8s objects like `pod`. Facilitates plugging in alternate Access Management.
- Make `userAccount` part of a separate API group from core K8s objects like
`pod.` Facilitates plugging in alternate Access Management.
Simple Profile:
- single `userAccount`, used by all K8s Users and Project Admins. One access token shared by all.
- Single `userAccount`, used by all K8s Users and Project Admins. One access
token shared by all.
Enterprise Profile:
- every human user has own `userAccount`.
- `userAccount`s have labels that indicate both membership in groups, and ability to act in certain roles.
- each service using the API has own `userAccount` too. (e.g. `scheduler`, `repcontroller`)
- automated jobs to denormalize the ldap group info into the local system list of users into the K8s userAccount file.
- Every human user has own `userAccount`.
- `userAccount`s have labels that indicate both membership in groups, and
ability to act in certain roles.
- Each service using the API has own `userAccount` too. (e.g. `scheduler`,
`repcontroller`)
- Automated jobs to denormalize the ldap group info into the local system
list of users into the K8s userAccount file.
### Unix accounts
A `userAccount` is not a Unix user account. The fact that a pod is started by a `userAccount` does not mean that the processes in that pod's containers run as a Unix user with a corresponding name or identity.
A `userAccount` is not a Unix user account. The fact that a pod is started by a
`userAccount` does not mean that the processes in that pod's containers run as a
Unix user with a corresponding name or identity.
Initially:
- The unix accounts available in a container, and used by the processes running in a container are those that are provided by the combination of the base operating system and the Docker manifest.
- Kubernetes doesn't enforce any relation between `userAccount` and unix accounts.
- The unix accounts available in a container, and used by the processes running
in a container are those that are provided by the combination of the base
operating system and the Docker manifest.
- Kubernetes doesn't enforce any relation between `userAccount` and unix
accounts.
Improvements:
- Kubelet allocates disjoint blocks of root-namespace uids for each container. This may provide some defense-in-depth against container escapes. (https://github.com/docker/docker/pull/4572)
- requires docker to integrate user namespace support, and deciding what getpwnam() does for these uids.
- any features that help users avoid use of privileged containers (http://issue.k8s.io/391)
- Kubelet allocates disjoint blocks of root-namespace uids for each container.
This may provide some defense-in-depth against container escapes. (https://github.com/docker/docker/pull/4572)
- requires docker to integrate user namespace support, and deciding what
getpwnam() does for these uids.
- any features that help users avoid use of privileged containers
(http://issue.k8s.io/391)
### Namespaces
K8s will have a `namespace` API object. It is similar to a Google Compute Engine `project`. It provides a namespace for objects created by a group of people co-operating together, preventing name collisions with non-cooperating groups. It also serves as a reference point for authorization policies.
K8s will have a `namespace` API object. It is similar to a Google Compute
Engine `project`. It provides a namespace for objects created by a group of
people co-operating together, preventing name collisions with non-cooperating
groups. It also serves as a reference point for authorization policies.
Namespaces are described in [namespaces.md](namespaces.md).
@ -192,20 +255,36 @@ In the Simple Profile:
- There is a single `namespace` used by the single user.
Namespaces versus userAccount vs Labels:
- `userAccount`s are intended for audit logging (both name and UID should be logged), and to define who has access to `namespace`s.
- `labels` (see [docs/user-guide/labels.md](../../docs/user-guide/labels.md)) should be used to distinguish pods, users, and other objects that cooperate towards a common goal but are different in some way, such as version, or responsibilities.
- `namespace`s prevent name collisions between uncoordinated groups of people, and provide a place to attach common policies for co-operating groups of people.
- `userAccount`s are intended for audit logging (both name and UID should be
logged), and to define who has access to `namespace`s.
- `labels` (see [docs/user-guide/labels.md](../../docs/user-guide/labels.md))
should be used to distinguish pods, users, and other objects that cooperate
towards a common goal but are different in some way, such as version, or
responsibilities.
- `namespace`s prevent name collisions between uncoordinated groups of people,
and provide a place to attach common policies for co-operating groups of people.
## Authentication
Goals for K8s authentication:
- Include a built-in authentication system with no configuration required to use in single-user mode, and little configuration required to add several user accounts, and no https proxy required.
- Allow for authentication to be handled by a system external to Kubernetes, to allow integration with existing to enterprise authorization systems. The Kubernetes namespace itself should avoid taking contributions of multiple authorization schemes. Instead, a trusted proxy in front of the apiserver can be used to authenticate users.
- For organizations whose security requirements only allow FIPS compliant implementations (e.g. apache) for authentication.
- So the proxy can terminate SSL, and isolate the CA-signed certificate from less trusted, higher-touch APIserver.
- For organizations that already have existing SaaS web services (e.g. storage, VMs) and want a common authentication portal.
- Avoid mixing authentication and authorization, so that authorization policies be centrally managed, and to allow changes in authentication methods without affecting authorization code.
- Include a built-in authentication system with no configuration required to use
in single-user mode, and little configuration required to add several user
accounts, and no https proxy required.
- Allow for authentication to be handled by a system external to Kubernetes, to
allow integration with existing to enterprise authorization systems. The
Kubernetes namespace itself should avoid taking contributions of multiple
authorization schemes. Instead, a trusted proxy in front of the apiserver can be
used to authenticate users.
- For organizations whose security requirements only allow FIPS compliant
implementations (e.g. apache) for authentication.
- So the proxy can terminate SSL, and isolate the CA-signed certificate from
less trusted, higher-touch APIserver.
- For organizations that already have existing SaaS web services (e.g.
storage, VMs) and want a common authentication portal.
- Avoid mixing authentication and authorization, so that authorization policies
be centrally managed, and to allow changes in authentication methods without
affecting authorization code.
Initially:
- Tokens used to authenticate a user.
@ -213,9 +292,12 @@ Initially:
- Administrator utility generates tokens at cluster setup.
- OAuth2.0 Bearer tokens protocol, http://tools.ietf.org/html/rfc6750
- No scopes for tokens. Authorization happens in the API server
- Tokens dynamically generated by apiserver to identify pods which are making API calls.
- Tokens dynamically generated by apiserver to identify pods which are making
API calls.
- Tokens checked in a module of the APIserver.
- Authentication in apiserver can be disabled by flag, to allow testing without authorization enabled, and to allow use of an authenticating proxy. In this mode, a query parameter or header added by the proxy will identify the caller.
- Authentication in apiserver can be disabled by flag, to allow testing without
authorization enabled, and to allow use of an authenticating proxy. In this
mode, a query parameter or header added by the proxy will identify the caller.
Improvements:
- Refresh of tokens.
@ -228,54 +310,86 @@ To be considered for subsequent versions:
- http://www.ietf.org/proceedings/90/slides/slides-90-uta-0.pdf
- http://www.browserauth.net
## Authorization
K8s authorization should:
- Allow for a range of maturity levels, from single-user for those test driving the system, to integration with existing to enterprise authorization systems.
- Allow for centralized management of users and policies. In some organizations, this will mean that the definition of users and access policies needs to reside on a system other than k8s and encompass other web services (such as a storage service).
- Allow processes running in K8s Pods to take on identity, and to allow narrow scoping of permissions for those identities in order to limit damage from software faults.
- Have Authorization Policies exposed as API objects so that a single config file can create or delete Pods, Replication Controllers, Services, and the identities and policies for those Pods and Replication Controllers.
- Be separate as much as practical from Authentication, to allow Authentication methods to change over time and space, without impacting Authorization policies.
- Allow for a range of maturity levels, from single-user for those test driving
the system, to integration with existing to enterprise authorization systems.
- Allow for centralized management of users and policies. In some
organizations, this will mean that the definition of users and access policies
needs to reside on a system other than k8s and encompass other web services
(such as a storage service).
- Allow processes running in K8s Pods to take on identity, and to allow narrow
scoping of permissions for those identities in order to limit damage from
software faults.
- Have Authorization Policies exposed as API objects so that a single config
file can create or delete Pods, Replication Controllers, Services, and the
identities and policies for those Pods and Replication Controllers.
- Be separate as much as practical from Authentication, to allow Authentication
methods to change over time and space, without impacting Authorization policies.
K8s will implement a relatively simple
[Attribute-Based Access Control](http://en.wikipedia.org/wiki/Attribute_Based_Access_Control) model.
The model will be described in more detail in a forthcoming document. The model will
The model will be described in more detail in a forthcoming document. The model
will:
- Be less complex than XACML
- Be easily recognizable to those familiar with Amazon IAM Policies.
- Have a subset/aliases/defaults which allow it to be used in a way comfortable to those users more familiar with Role-Based Access Control.
- Have a subset/aliases/defaults which allow it to be used in a way comfortable
to those users more familiar with Role-Based Access Control.
Authorization policy is set by creating a set of Policy objects.
The API Server will be the Enforcement Point for Policy. For each API call that it receives, it will construct the Attributes needed to evaluate the policy (what user is making the call, what resource they are accessing, what they are trying to do that resource, etc) and pass those attributes to a Decision Point. The Decision Point code evaluates the Attributes against all the Policies and allows or denies the API call. The system will be modular enough that the Decision Point code can either be linked into the APIserver binary, or be another service that the apiserver calls for each Decision (with appropriate time-limited caching as needed for performance).
Policy objects may be applicable only to a single namespace or to all namespaces; K8s Project Admins would be able to create those as needed. Other Policy objects may be applicable to all namespaces; a K8s Cluster Admin might create those in order to authorize a new type of controller to be used by all namespaces, or to make a K8s User into a K8s Project Admin.)
The API Server will be the Enforcement Point for Policy. For each API call that
it receives, it will construct the Attributes needed to evaluate the policy
(what user is making the call, what resource they are accessing, what they are
trying to do that resource, etc) and pass those attributes to a Decision Point.
The Decision Point code evaluates the Attributes against all the Policies and
allows or denies the API call. The system will be modular enough that the
Decision Point code can either be linked into the APIserver binary, or be
another service that the apiserver calls for each Decision (with appropriate
time-limited caching as needed for performance).
Policy objects may be applicable only to a single namespace or to all
namespaces; K8s Project Admins would be able to create those as needed. Other
Policy objects may be applicable to all namespaces; a K8s Cluster Admin might
create those in order to authorize a new type of controller to be used by all
namespaces, or to make a K8s User into a K8s Project Admin.)
## Accounting
The API should have a `quota` concept (see http://issue.k8s.io/442). A quota object relates a namespace (and optionally a label selector) to a maximum quantity of resources that may be used (see [resources design doc](resources.md)).
The API should have a `quota` concept (see http://issue.k8s.io/442). A quota
object relates a namespace (and optionally a label selector) to a maximum
quantity of resources that may be used (see [resources design doc](resources.md)).
Initially:
- a `quota` object is immutable.
- for hosted K8s systems that do billing, Project is recommended level for billing accounts.
- Every object that consumes resources should have a `namespace` so that Resource usage stats are roll-up-able to `namespace`.
- A `quota` object is immutable.
- For hosted K8s systems that do billing, Project is recommended level for
billing accounts.
- Every object that consumes resources should have a `namespace` so that
Resource usage stats are roll-up-able to `namespace`.
- K8s Cluster Admin sets quota objects by writing a config file.
Improvements:
- allow one namespace to charge the quota for one or more other namespaces. This would be controlled by a policy which allows changing a billing_namespace= label on an object.
- allow quota to be set by namespace owners for (namespace x label) combinations (e.g. let "webserver" namespace use 100 cores, but to prevent accidents, don't allow "webserver" namespace and "instance=test" use more than 10 cores.
- tools to help write consistent quota config files based on number of nodes, historical namespace usages, QoS needs, etc.
- way for K8s Cluster Admin to incrementally adjust Quota objects.
- Allow one namespace to charge the quota for one or more other namespaces. This
would be controlled by a policy which allows changing a billing_namespace =
label on an object.
- Allow quota to be set by namespace owners for (namespace x label) combinations
(e.g. let "webserver" namespace use 100 cores, but to prevent accidents, don't
allow "webserver" namespace and "instance=test" use more than 10 cores.
- Tools to help write consistent quota config files based on number of nodes,
historical namespace usages, QoS needs, etc.
- Way for K8s Cluster Admin to incrementally adjust Quota objects.
Simple profile:
- a single `namespace` with infinite resource limits.
- A single `namespace` with infinite resource limits.
Enterprise profile:
- multiple namespaces each with their own limits.
- Multiple namespaces each with their own limits.
Issues:
- need for locking or "eventual consistency" when multiple apiserver goroutines are accessing the object store and handling pod creations.
- Need for locking or "eventual consistency" when multiple apiserver goroutines
are accessing the object store and handling pod creations.
## Audit Logging
@ -287,7 +401,8 @@ Initial implementation:
Improvements:
- API server does logging instead.
- Policies to drop logging for high rate trusted API calls, or by users performing audit or other sensitive functions.
- Policies to drop logging for high rate trusted API calls, or by users
performing audit or other sensitive functions.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

38
docs/design/admission_control.md
View File

@ -43,24 +43,30 @@ Documentation for other releases can be found at
## Background
High level goals:
* Enable an easy-to-use mechanism to provide admission control to cluster.
* Enable a provider to support multiple admission control strategies or author
their own.
* Ensure any rejected request can propagate errors back to the caller with why
the request failed.
* Enable an easy-to-use mechanism to provide admission control to cluster
* Enable a provider to support multiple admission control strategies or author their own
* Ensure any rejected request can propagate errors back to the caller with why the request failed
Authorization via policy is focused on answering if a user is authorized to perform an action.
Authorization via policy is focused on answering if a user is authorized to
perform an action.
Admission Control is focused on if the system will accept an authorized action.
Kubernetes may choose to dismiss an authorized action based on any number of admission control strategies.
Kubernetes may choose to dismiss an authorized action based on any number of
admission control strategies.
This proposal documents the basic design, and describes how any number of admission control plug-ins could be injected.
This proposal documents the basic design, and describes how any number of
admission control plug-ins could be injected.
Implementation of specific admission control strategies are handled in separate documents.
Implementation of specific admission control strategies are handled in separate
documents.
## kube-apiserver
The kube-apiserver takes the following OPTIONAL arguments to enable admission control
The kube-apiserver takes the following OPTIONAL arguments to enable admission
control:
| Option | Behavior |
| ------ | -------- |
@ -72,7 +78,8 @@ An **AdmissionControl** plug-in is an implementation of the following interface:
```go
package admission
// Attributes is an interface used by a plug-in to make an admission decision on a individual request.
// Attributes is an interface used by a plug-in to make an admission decision
// on a individual request.
type Attributes interface {
GetNamespace() string
GetKind() string
@ -88,8 +95,8 @@ type Interface interface {
}
```
A **plug-in** must be compiled with the binary, and is registered as an available option by providing a name, and implementation
of admission.Interface.
A **plug-in** must be compiled with the binary, and is registered as an
available option by providing a name, and implementation of admission.Interface.
```go
func init() {
@ -97,9 +104,12 @@ func init() {
}
```
Invocation of admission control is handled by the **APIServer** and not individual **RESTStorage** implementations.
Invocation of admission control is handled by the **APIServer** and not
individual **RESTStorage** implementations.
This design assumes that **Issue 297** is adopted, and as a consequence, the general framework of the APIServer request/response flow will ensure the following:
This design assumes that **Issue 297** is adopted, and as a consequence, the
general framework of the APIServer request/response flow will ensure the
following:
1. Incoming request
2. Authenticate user

67
docs/design/admission_control_limit_range.md
View File

@ -36,7 +36,8 @@ Documentation for other releases can be found at
## Background
This document proposes a system for enforcing resource requirements constraints as part of admission control.
This document proposes a system for enforcing resource requirements constraints
as part of admission control.
## Use cases
@ -64,7 +65,8 @@ const (
LimitTypeContainer LimitType = "Container"
)
// LimitRangeItem defines a min/max usage limit for any resource that matches on kind.
// LimitRangeItem defines a min/max usage limit for any resource that matches
// on kind.
type LimitRangeItem struct {
// Type of resource that this limit applies to.
Type LimitType `json:"type,omitempty"`
@ -72,29 +74,38 @@ type LimitRangeItem struct {
Max ResourceList `json:"max,omitempty"`
// Min usage constraints on this kind by resource name.
Min ResourceList `json:"min,omitempty"`
// Default resource requirement limit value by resource name if resource limit is omitted.
// Default resource requirement limit value by resource name if resource limit
// is omitted.
Default ResourceList `json:"default,omitempty"`
// DefaultRequest is the default resource requirement request value by resource name if resource request is omitted.
// DefaultRequest is the default resource requirement request value by
// resource name if resource request is omitted.
DefaultRequest ResourceList `json:"defaultRequest,omitempty"`
// MaxLimitRequestRatio if specified, the named resource must have a request and limit that are both non-zero where limit divided by request is less than or equal to the enumerated value; this represents the max burst for the named resource.
// MaxLimitRequestRatio if specified, the named resource must have a request
// and limit that are both non-zero where limit divided by request is less
// than or equal to the enumerated value; this represents the max burst for
// the named resource.
MaxLimitRequestRatio ResourceList `json:"maxLimitRequestRatio,omitempty"`
}
// LimitRangeSpec defines a min/max usage limit for resources that match on kind.
// LimitRangeSpec defines a min/max usage limit for resources that match
// on kind.
type LimitRangeSpec struct {
// Limits is the list of LimitRangeItem objects that are enforced.
Limits []LimitRangeItem `json:"limits"`
}
// LimitRange sets resource usage limits for each kind of resource in a Namespace.
// LimitRange sets resource usage limits for each kind of resource in a
// Namespace.
type LimitRange struct {
TypeMeta `json:",inline"`
// Standard object's metadata.
// More info: http://releases.k8s.io/HEAD/docs/devel/api-conventions.md#metadata
// More info:
// http://releases.k8s.io/HEAD/docs/devel/api-conventions.md#metadata
ObjectMeta `json:"metadata,omitempty"`
// Spec defines the limits enforced.
// More info: http://releases.k8s.io/HEAD/docs/devel/api-conventions.md#spec-and-status
// More info:
// http://releases.k8s.io/HEAD/docs/devel/api-conventions.md#spec-and-status
Spec LimitRangeSpec `json:"spec,omitempty"`
}
@ -102,24 +113,29 @@ type LimitRange struct {
type LimitRangeList struct {
TypeMeta `json:",inline"`
// Standard list metadata.
// More info: http://releases.k8s.io/HEAD/docs/devel/api-conventions.md#types-kinds
// More info:
// http://releases.k8s.io/HEAD/docs/devel/api-conventions.md#types-kinds
ListMeta `json:"metadata,omitempty"`
// Items is a list of LimitRange objects.
// More info: http://releases.k8s.io/HEAD/docs/design/admission_control_limit_range.md
// More info:
// http://releases.k8s.io/HEAD/docs/design/admission_control_limit_range.md
Items []LimitRange `json:"items"`
}
```
### Validation
Validation of a **LimitRange** enforces that for a given named resource the following rules apply:
Validation of a **LimitRange** enforces that for a given named resource the
following rules apply:
Min (if specified) <= DefaultRequest (if specified) <= Default (if specified) <= Max (if specified)
Min (if specified) <= DefaultRequest (if specified) <= Default (if specified)
<= Max (if specified)
### Default Value Behavior
The following default value behaviors are applied to a LimitRange for a given named resource.
The following default value behaviors are applied to a LimitRange for a given
named resource.
```
if LimitRangeItem.Default[resourceName] is undefined
@ -137,11 +153,14 @@ if LimitRangeItem.DefaultRequest[resourceName] is undefined
## AdmissionControl plugin: LimitRanger
The **LimitRanger** plug-in introspects all incoming pod requests and evaluates the constraints defined on a LimitRange.
The **LimitRanger** plug-in introspects all incoming pod requests and evaluates
the constraints defined on a LimitRange.
If a constraint is not specified for an enumerated resource, it is not enforced or tracked.
If a constraint is not specified for an enumerated resource, it is not enforced
or tracked.
To enable the plug-in and support for LimitRange, the kube-apiserver must be configured as follows:
To enable the plug-in and support for LimitRange, the kube-apiserver must be
configured as follows:
```console
$ kube-apiserver --admission-control=LimitRanger
@ -158,7 +177,7 @@ Supported Resources:
Supported Constraints:
Per container, the following must hold true
Per container, the following must hold true:
| Constraint | Behavior |
| ---------- | -------- |
@ -168,8 +187,10 @@ Per container, the following must hold true
Supported Defaults:
1. Default - if the named resource has no enumerated value, the Limit is equal to the Default
2. DefaultRequest - if the named resource has no enumerated value, the Request is equal to the DefaultRequest
1. Default - if the named resource has no enumerated value, the Limit is equal
to the Default
2. DefaultRequest - if the named resource has no enumerated value, the Request
is equal to the DefaultRequest
**Type: Pod**
@ -190,7 +211,8 @@ Across all containers in pod, the following must hold true
## Run-time configuration
The default ```LimitRange``` that is applied via Salt configuration will be updated as follows:
The default ```LimitRange``` that is applied via Salt configuration will be
updated as follows:
```
apiVersion: "v1"
@ -219,7 +241,8 @@ the following would happen.
1. The incoming container cpu would request 250m with a limit of 500m.
2. The incoming container memory would request 250Mi with a limit of 500Mi
3. If the container is later resized, it's cpu would be constrained to between .1 and 1 and the ratio of limit to request could not exceed 4.
3. If the container is later resized, it's cpu would be constrained to between
.1 and 1 and the ratio of limit to request could not exceed 4.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->
[![Analytics](https://kubernetes-site.appspot.com/UA-36037335-10/GitHub/docs/design/admission_control_limit_range.md?pixel)]()

91
docs/design/admission_control_resource_quota.md
View File

@ -36,7 +36,8 @@ Documentation for other releases can be found at
## Background
This document describes a system for enforcing hard resource usage limits per namespace as part of admission control.
This document describes a system for enforcing hard resource usage limits per
namespace as part of admission control.
## Use cases
@ -103,7 +104,7 @@ type ResourceQuotaList struct {
## Quota Tracked Resources
The following resources are supported by the quota system.
The following resources are supported by the quota system:
| Resource | Description |
| ------------ | ----------- |
@ -116,16 +117,19 @@ The following resources are supported by the quota system.
| secrets | Total number of secrets |
| persistentvolumeclaims | Total number of persistent volume claims |
If a third-party wants to track additional resources, it must follow the resource naming conventions prescribed
by Kubernetes. This means the resource must have a fully-qualified name (i.e. mycompany.org/shinynewresource)
If a third-party wants to track additional resources, it must follow the
resource naming conventions prescribed by Kubernetes. This means the resource
must have a fully-qualified name (i.e. mycompany.org/shinynewresource)
## Resource Requirements: Requests vs Limits
If a resource supports the ability to distinguish between a request and a limit for a resource,
the quota tracking system will only cost the request value against the quota usage. If a resource
is tracked by quota, and no request value is provided, the associated entity is rejected as part of admission.
If a resource supports the ability to distinguish between a request and a limit
for a resource, the quota tracking system will only cost the request value
against the quota usage. If a resource is tracked by quota, and no request value
is provided, the associated entity is rejected as part of admission.
For an example, consider the following scenarios relative to tracking quota on CPU:
For an example, consider the following scenarios relative to tracking quota on
CPU:
| Pod | Container | Request CPU | Limit CPU | Result |
| --- | --------- | ----------- | --------- | ------ |
@ -134,13 +138,14 @@ For an example, consider the following scenarios relative to tracking quota on C
| Y | C2 | none | 500m | The quota usage is incremented 500m since request will default to limit |
| Z | C3 | none | none | The pod is rejected since it does not enumerate a request. |
The rationale for accounting for the requested amount of a resource versus the limit is the belief
that a user should only be charged for what they are scheduled against in the cluster. In addition,
attempting to track usage against actual usage, where request < actual < limit, is considered highly
volatile.
The rationale for accounting for the requested amount of a resource versus the
limit is the belief that a user should only be charged for what they are
scheduled against in the cluster. In addition, attempting to track usage against
actual usage, where request < actual < limit, is considered highly volatile.
As a consequence of this decision, the user is able to spread its usage of a resource across multiple tiers
of service. Let's demonstrate this via an example with a 4 cpu quota.
As a consequence of this decision, the user is able to spread its usage of a
resource across multiple tiers of service. Let's demonstrate this via an
example with a 4 cpu quota.
The quota may be allocated as follows:
@ -150,48 +155,62 @@ The quota may be allocated as follows:
| Y | C2 | 2 | 2 | Guaranteed | 2 |
| Z | C3 | 1 | 3 | Burstable | 1 |
It is possible that the pods may consume 9 cpu over a given time period depending on the nodes available cpu
that held pod X and Z, but since we scheduled X and Z relative to the request, we only track the requesting
value against their allocated quota. If one wants to restrict the ratio between the request and limit,
it is encouraged that the user define a **LimitRange** with **LimitRequestRatio** to control burst out behavior.
This would in effect, let an administrator keep the difference between request and limit more in line with
It is possible that the pods may consume 9 cpu over a given time period
depending on the nodes available cpu that held pod X and Z, but since we
scheduled X and Z relative to the request, we only track the requesting value
against their allocated quota. If one wants to restrict the ratio between the
request and limit, it is encouraged that the user define a **LimitRange** with
**LimitRequestRatio** to control burst out behavior. This would in effect, let
an administrator keep the difference between request and limit more in line with
tracked usage if desired.
## Status API
A REST API endpoint to update the status section of the **ResourceQuota** is exposed. It requires an atomic compare-and-swap
in order to keep resource usage tracking consistent.
A REST API endpoint to update the status section of the **ResourceQuota** is
exposed. It requires an atomic compare-and-swap in order to keep resource usage
tracking consistent.
## Resource Quota Controller
A resource quota controller monitors observed usage for tracked resources in the **Namespace**.
A resource quota controller monitors observed usage for tracked resources in the
**Namespace**.
If there is observed difference between the current usage stats versus the current **ResourceQuota.Status**, the controller
posts an update of the currently observed usage metrics to the **ResourceQuota** via the /status endpoint.
If there is observed difference between the current usage stats versus the
current **ResourceQuota.Status**, the controller posts an update of the
currently observed usage metrics to the **ResourceQuota** via the /status
endpoint.
The resource quota controller is the only component capable of monitoring and recording usage updates after a DELETE operation
since admission control is incapable of guaranteeing a DELETE request actually succeeded.
The resource quota controller is the only component capable of monitoring and
recording usage updates after a DELETE operation since admission control is
incapable of guaranteeing a DELETE request actually succeeded.
## AdmissionControl plugin: ResourceQuota
The **ResourceQuota** plug-in introspects all incoming admission requests.
To enable the plug-in and support for ResourceQuota, the kube-apiserver must be configured as follows:
To enable the plug-in and support for ResourceQuota, the kube-apiserver must be
configured as follows:
```
$ kube-apiserver --admission-control=ResourceQuota
```
It makes decisions by evaluating the incoming object against all defined **ResourceQuota.Status.Hard** resource limits in the request
namespace. If acceptance of the resource would cause the total usage of a named resource to exceed its hard limit, the request is denied.
It makes decisions by evaluating the incoming object against all defined
**ResourceQuota.Status.Hard** resource limits in the request namespace. If
acceptance of the resource would cause the total usage of a named resource to
exceed its hard limit, the request is denied.
If the incoming request does not cause the total usage to exceed any of the enumerated hard resource limits, the plug-in will post a
**ResourceQuota.Status** document to the server to atomically update the observed usage based on the previously read
**ResourceQuota.ResourceVersion**. This keeps incremental usage atomically consistent, but does introduce a bottleneck (intentionally)
into the system.
If the incoming request does not cause the total usage to exceed any of the
enumerated hard resource limits, the plug-in will post a
**ResourceQuota.Status** document to the server to atomically update the
observed usage based on the previously read **ResourceQuota.ResourceVersion**.
This keeps incremental usage atomically consistent, but does introduce a
bottleneck (intentionally) into the system.
To optimize system performance, it is encouraged that all resource quotas are tracked on the same **ResourceQuota** document in a **Namespace**. As a result, its encouraged to impose a cap on the total number of individual quotas that are tracked in the **Namespace**
to 1 in the **ResourceQuota** document.
To optimize system performance, it is encouraged that all resource quotas are
tracked on the same **ResourceQuota** document in a **Namespace**. As a result,
it is encouraged to impose a cap on the total number of individual quotas that
are tracked in the **Namespace** to 1 in the **ResourceQuota** document.
## kubectl
@ -199,7 +218,7 @@ kubectl is modified to support the **ResourceQuota** resource.
`kubectl describe` provides a human-readable output of quota.
For example,
For example:
```console
$ kubectl create -f docs/admin/resourcequota/namespace.yaml

61
docs/design/architecture.md
View File

@ -34,49 +34,84 @@ Documentation for other releases can be found at
# Kubernetes architecture
A running Kubernetes cluster contains node agents (`kubelet`) and master components (APIs, scheduler, etc), on top of a distributed storage solution. This diagram shows our desired eventual state, though we're still working on a few things, like making `kubelet` itself (all our components, really) run within containers, and making the scheduler 100% pluggable.
A running Kubernetes cluster contains node agents (`kubelet`) and master
components (APIs, scheduler, etc), on top of a distributed storage solution.
This diagram shows our desired eventual state, though we're still working on a
few things, like making `kubelet` itself (all our components, really) run within
containers, and making the scheduler 100% pluggable.
![Architecture Diagram](architecture.png?raw=true "Architecture overview")
## The Kubernetes Node
When looking at the architecture of the system, we'll break it down to services that run on the worker node and services that compose the cluster-level control plane.
When looking at the architecture of the system, we'll break it down to services
that run on the worker node and services that compose the cluster-level control
plane.
The Kubernetes node has the services necessary to run application containers and be managed from the master systems.
The Kubernetes node has the services necessary to run application containers and
be managed from the master systems.
Each node runs Docker, of course. Docker takes care of the details of downloading images and running containers.
Each node runs Docker, of course. Docker takes care of the details of
downloading images and running containers.
### `kubelet`
The `kubelet` manages [pods](../user-guide/pods.md) and their containers, their images, their volumes, etc.
The `kubelet` manages [pods](../user-guide/pods.md) and their containers, their
images, their volumes, etc.
### `kube-proxy`
Each node also runs a simple network proxy and load balancer (see the [services FAQ](https://github.com/kubernetes/kubernetes/wiki/Services-FAQ) for more details). This reflects `services` (see [the services doc](../user-guide/services.md) for more details) as defined in the Kubernetes API on each node and can do simple TCP and UDP stream forwarding (round robin) across a set of backends.
Each node also runs a simple network proxy and load balancer (see the
[services FAQ](https://github.com/kubernetes/kubernetes/wiki/Services-FAQ) for
more details). This reflects `services` (see
[the services doc](../user-guide/services.md) for more details) as defined in
the Kubernetes API on each node and can do simple TCP and UDP stream forwarding
(round robin) across a set of backends.
Service endpoints are currently found via [DNS](../admin/dns.md) or through environment variables (both [Docker-links-compatible](https://docs.docker.com/userguide/dockerlinks/) and Kubernetes `{FOO}_SERVICE_HOST` and `{FOO}_SERVICE_PORT` variables are supported). These variables resolve to ports managed by the service proxy.
Service endpoints are currently found via [DNS](../admin/dns.md) or through
environment variables (both
[Docker-links-compatible](https://docs.docker.com/userguide/dockerlinks/) and
Kubernetes `{FOO}_SERVICE_HOST` and `{FOO}_SERVICE_PORT` variables are
supported). These variables resolve to ports managed by the service proxy.
## The Kubernetes Control Plane
The Kubernetes control plane is split into a set of components. Currently they all run on a single _master_ node, but that is expected to change soon in order to support high-availability clusters. These components work together to provide a unified view of the cluster.
The Kubernetes control plane is split into a set of components. Currently they
all run on a single _master_ node, but that is expected to change soon in order
to support high-availability clusters. These components work together to provide
a unified view of the cluster.
### `etcd`
All persistent master state is stored in an instance of `etcd`. This provides a great way to store configuration data reliably. With `watch` support, coordinating components can be notified very quickly of changes.
All persistent master state is stored in an instance of `etcd`. This provides a
great way to store configuration data reliably. With `watch` support,
coordinating components can be notified very quickly of changes.
### Kubernetes API Server
The apiserver serves up the [Kubernetes API](../api.md). It is intended to be a CRUD-y server, with most/all business logic implemented in separate components or in plug-ins. It mainly processes REST operations, validates them, and updates the corresponding objects in `etcd` (and eventually other stores).
The apiserver serves up the [Kubernetes API](../api.md). It is intended to be a
CRUD-y server, with most/all business logic implemented in separate components
or in plug-ins. It mainly processes REST operations, validates them, and updates
the corresponding objects in `etcd` (and eventually other stores).
### Scheduler
The scheduler binds unscheduled pods to nodes via the `/binding` API. The scheduler is pluggable, and we expect to support multiple cluster schedulers and even user-provided schedulers in the future.
The scheduler binds unscheduled pods to nodes via the `/binding` API. The
scheduler is pluggable, and we expect to support multiple cluster schedulers and
even user-provided schedulers in the future.
### Kubernetes Controller Manager Server
All other cluster-level functions are currently performed by the Controller Manager. For instance, `Endpoints` objects are created and updated by the endpoints controller, and nodes are discovered, managed, and monitored by the node controller. These could eventually be split into separate components to make them independently pluggable.
All other cluster-level functions are currently performed by the Controller
Manager. For instance, `Endpoints` objects are created and updated by the
endpoints controller, and nodes are discovered, managed, and monitored by the
node controller. These could eventually be split into separate components to
make them independently pluggable.
The [`replicationcontroller`](../user-guide/replication-controller.md) is a mechanism that is layered on top of the simple [`pod`](../user-guide/pods.md) API. We eventually plan to port it to a generic plug-in mechanism, once one is implemented.
The [`replicationcontroller`](../user-guide/replication-controller.md) is a
mechanism that is layered on top of the simple [`pod`](../user-guide/pods.md)
API. We eventually plan to port it to a generic plug-in mechanism, once one is
implemented.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

180
docs/design/aws_under_the_hood.md
View File

@ -35,7 +35,7 @@ Documentation for other releases can be found at
# Peeking under the hood of Kubernetes on AWS
This document provides high-level insight into how Kubernetes works on AWS and
maps to AWS objects. We assume that you are familiar with AWS.
maps to AWS objects. We assume that you are familiar with AWS.
We encourage you to use [kube-up](../getting-started-guides/aws.md) to create
clusters on AWS. We recommend that you avoid manual configuration but are aware
@ -72,7 +72,7 @@ By default on AWS:
* Instances run Ubuntu 15.04 (the official AMI). It includes a sufficiently
modern kernel that pairs well with Docker and doesn't require a
reboot. (The default SSH user is `ubuntu` for this and other ubuntu images.)
reboot. (The default SSH user is `ubuntu` for this and other ubuntu images.)
* Nodes use aufs instead of ext4 as the filesystem / container storage (mostly
because this is what Google Compute Engine uses).
@ -81,35 +81,36 @@ kube-up.
### Storage
AWS supports persistent volumes by using [Elastic Block Store (EBS)](../user-guide/volumes.md#awselasticblockstore). These can then be
attached to pods that should store persistent data (e.g. if you're running a
database).
AWS supports persistent volumes by using [Elastic Block Store (EBS)](../user-guide/volumes.md#awselasticblockstore).
These can then be attached to pods that should store persistent data (e.g. if
you're running a database).
By default, nodes in AWS use [instance storage](http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/InstanceStorage.html)
unless you create pods with persistent volumes
[(EBS)](../user-guide/volumes.md#awselasticblockstore). In general, Kubernetes
[(EBS)](../user-guide/volumes.md#awselasticblockstore). In general, Kubernetes
containers do not have persistent storage unless you attach a persistent
volume, and so nodes on AWS use instance storage. Instance storage is cheaper,
often faster, and historically more reliable. Unless you can make do with whatever
space is left on your root partition, you must choose an instance type that provides
you with sufficient instance storage for your needs.
often faster, and historically more reliable. Unless you can make do with
whatever space is left on your root partition, you must choose an instance type
that provides you with sufficient instance storage for your needs.
Note: The master uses a persistent volume ([etcd](architecture.md#etcd)) to track
its state. Similar to nodes, containers are mostly run against instance
Note: The master uses a persistent volume ([etcd](architecture.md#etcd)) to
track its state. Similar to nodes, containers are mostly run against instance
storage, except that we repoint some important data onto the persistent volume.
The default storage driver for Docker images is aufs. Specifying btrfs (by passing the environment
variable `DOCKER_STORAGE=btrfs` to kube-up) is also a good choice for a filesystem. btrfs
is relatively reliable with Docker and has improved its reliability with modern
kernels. It can easily span multiple volumes, which is particularly useful
when we are using an instance type with multiple ephemeral instance disks.
The default storage driver for Docker images is aufs. Specifying btrfs (by
passing the environment variable `DOCKER_STORAGE=btrfs` to kube-up) is also a
good choice for a filesystem. btrfs is relatively reliable with Docker and has
improved its reliability with modern kernels. It can easily span multiple
volumes, which is particularly useful when we are using an instance type with
multiple ephemeral instance disks.
### Auto Scaling group
Nodes (but not the master) are run in an
[Auto Scaling group](http://docs.aws.amazon.com/AutoScaling/latest/DeveloperGuide/AutoScalingGroup.html)
on AWS. Currently auto-scaling (e.g. based on CPU) is not actually enabled
([#11935](http://issues.k8s.io/11935)). Instead, the Auto Scaling group means
on AWS. Currently auto-scaling (e.g. based on CPU) is not actually enabled
([#11935](http://issues.k8s.io/11935)). Instead, the Auto Scaling group means
that AWS will relaunch any nodes that are terminated.
We do not currently run the master in an AutoScalingGroup, but we should
@ -117,13 +118,13 @@ We do not currently run the master in an AutoScalingGroup, but we should
### Networking
Kubernetes uses an IP-per-pod model. This means that a node, which runs many
pods, must have many IPs. AWS uses virtual private clouds (VPCs) and advanced
Kubernetes uses an IP-per-pod model. This means that a node, which runs many
pods, must have many IPs. AWS uses virtual private clouds (VPCs) and advanced
routing support so each pod is assigned a /24 CIDR. The assigned CIDR is then
configured to route to an instance in the VPC routing table.
It is also possible to use overlay networking on AWS, but that is not the default
configuration of the kube-up script.
It is also possible to use overlay networking on AWS, but that is not the
default configuration of the kube-up script.
### NodePort and LoadBalancer services
@ -137,8 +138,8 @@ the nodes. This traffic reaches kube-proxy where it is then forwarded to the
pods.
ELB has some restrictions:
* it requires that all nodes listen on a single port,
* it acts as a forwarding proxy (i.e. the source IP is not preserved).
* ELB requires that all nodes listen on a single port,
* ELB acts as a forwarding proxy (i.e. the source IP is not preserved).
To work with these restrictions, in Kubernetes, [LoadBalancer
services](../user-guide/services.md#type-loadbalancer) are exposed as
@ -146,18 +147,18 @@ services](../user-guide/services.md#type-loadbalancer) are exposed as
kube-proxy listens externally on the cluster-wide port that's assigned to
NodePort services and forwards traffic to the corresponding pods.
So for example, if we configure a service of Type LoadBalancer with a
For example, if we configure a service of Type LoadBalancer with a
public port of 80:
* Kubernetes will assign a NodePort to the service (e.g. 31234)
* Kubernetes will assign a NodePort to the service (e.g. port 31234)
* ELB is configured to proxy traffic on the public port 80 to the NodePort
that is assigned to the service (31234).
* Then any in-coming traffic that ELB forwards to the NodePort (e.g. port 31234)
is recognized by kube-proxy and sent to the correct pods for that service.
assigned to the service (in this example port 31234).
* Then any in-coming traffic that ELB forwards to the NodePort (31234)
is recognized by kube-proxy and sent to the correct pods for that service.
Note that we do not automatically open NodePort services in the AWS firewall
(although we do open LoadBalancer services). This is because we expect that
(although we do open LoadBalancer services). This is because we expect that
NodePort services are more of a building block for things like inter-cluster
services or for LoadBalancer. To consume a NodePort service externally, you
services or for LoadBalancer. To consume a NodePort service externally, you
will likely have to open the port in the node security group
(`kubernetes-minion-<clusterid>`).
@ -169,19 +170,19 @@ and one for the nodes called
[kubernetes-minion](../../cluster/aws/templates/iam/kubernetes-minion-policy.json).
The master is responsible for creating ELBs and configuring them, as well as
setting up advanced VPC routing. Currently it has blanket permissions on EC2,
setting up advanced VPC routing. Currently it has blanket permissions on EC2,
along with rights to create and destroy ELBs.
The nodes do not need a lot of access to the AWS APIs. They need to download
The nodes do not need a lot of access to the AWS APIs. They need to download
a distribution file, and then are responsible for attaching and detaching EBS
volumes from itself.
The node policy is relatively minimal. In 1.2 and later, nodes can retrieve ECR
The node policy is relatively minimal. In 1.2 and later, nodes can retrieve ECR
authorization tokens, refresh them every 12 hours if needed, and fetch Docker
images from it, as long as the appropriate permissions are enabled. Those in
images from it, as long as the appropriate permissions are enabled. Those in
[AmazonEC2ContainerRegistryReadOnly](http://docs.aws.amazon.com/AmazonECR/latest/userguide/ecr_managed_policies.html#AmazonEC2ContainerRegistryReadOnly),
without write access, should suffice. The master policy is probably overly
permissive. The security conscious may want to lock-down the IAM policies
without write access, should suffice. The master policy is probably overly
permissive. The security conscious may want to lock-down the IAM policies
further ([#11936](http://issues.k8s.io/11936)).
We should make it easier to extend IAM permissions and also ensure that they
@ -190,106 +191,101 @@ are correctly configured ([#14226](http://issues.k8s.io/14226)).
### Tagging
All AWS resources are tagged with a tag named "KubernetesCluster", with a value
that is the unique cluster-id. This tag is used to identify a particular
that is the unique cluster-id. This tag is used to identify a particular
'instance' of Kubernetes, even if two clusters are deployed into the same VPC.
Resources are considered to belong to the same cluster if and only if they have
the same value in the tag named "KubernetesCluster". (The kube-up script is
the same value in the tag named "KubernetesCluster". (The kube-up script is
not configured to create multiple clusters in the same VPC by default, but it
is possible to create another cluster in the same VPC.)
Within the AWS cloud provider logic, we filter requests to the AWS APIs to
match resources with our cluster tag. By filtering the requests, we ensure
match resources with our cluster tag. By filtering the requests, we ensure
that we see only our own AWS objects.
Important: If you choose not to use kube-up, you must pick a unique cluster-id
value, and ensure that all AWS resources have a tag with
** Important: ** If you choose not to use kube-up, you must pick a unique
cluster-id value, and ensure that all AWS resources have a tag with
`Name=KubernetesCluster,Value=<clusterid>`.
### AWS objects
The kube-up script does a number of things in AWS:
* Creates an S3 bucket (`AWS_S3_BUCKET`) and then copies the Kubernetes distribution
and the salt scripts into it. They are made world-readable and the HTTP URLs
are passed to instances; this is how Kubernetes code gets onto the machines.
* Creates an S3 bucket (`AWS_S3_BUCKET`) and then copies the Kubernetes
distribution and the salt scripts into it. They are made world-readable and the
HTTP URLs are passed to instances; this is how Kubernetes code gets onto the
machines.
* Creates two IAM profiles based on templates in [cluster/aws/templates/iam](../../cluster/aws/templates/iam/):
* `kubernetes-master` is used by the master.
* `kubernetes-minion` is used by nodes.
* Creates an AWS SSH key named `kubernetes-<fingerprint>`. Fingerprint here is
the OpenSSH key fingerprint, so that multiple users can run the script with
different keys and their keys will not collide (with near-certainty). It will
use an existing key if one is found at `AWS_SSH_KEY`, otherwise it will create
one there. (With the default Ubuntu images, if you have to SSH in: the user is
`ubuntu` and that user can `sudo`).
* Creates an AWS SSH key named `kubernetes-<fingerprint>`. Fingerprint here is
the OpenSSH key fingerprint, so that multiple users can run the script with
different keys and their keys will not collide (with near-certainty). It will
use an existing key if one is found at `AWS_SSH_KEY`, otherwise it will create
one there. (With the default Ubuntu images, if you have to SSH in: the user is
`ubuntu` and that user can `sudo`).
* Creates a VPC for use with the cluster (with a CIDR of 172.20.0.0/16) and
enables the `dns-support` and `dns-hostnames` options.
enables the `dns-support` and `dns-hostnames` options.
* Creates an internet gateway for the VPC.
* Creates a route table for the VPC, with the internet gateway as the default
route.
route.
* Creates a subnet (with a CIDR of 172.20.0.0/24) in the AZ `KUBE_AWS_ZONE`
(defaults to us-west-2a). Currently, each Kubernetes cluster runs in a
single AZ on AWS. Although, there are two philosophies in discussion on how to
achieve High Availability (HA):
* cluster-per-AZ: An independent cluster for each AZ, where each cluster
is entirely separate.
* cross-AZ-clusters: A single cluster spans multiple AZs.
(defaults to us-west-2a). Currently, each Kubernetes cluster runs in a
single AZ on AWS. Although, there are two philosophies in discussion on how to
achieve High Availability (HA):
* cluster-per-AZ: An independent cluster for each AZ, where each cluster
is entirely separate.
* cross-AZ-clusters: A single cluster spans multiple AZs.
The debate is open here, where cluster-per-AZ is discussed as more robust but
cross-AZ-clusters are more convenient.
* Associates the subnet to the route table
* Creates security groups for the master (`kubernetes-master-<clusterid>`)
and the nodes (`kubernetes-minion-<clusterid>`).
and the nodes (`kubernetes-minion-<clusterid>`).
* Configures security groups so that masters and nodes can communicate. This
includes intercommunication between masters and nodes, opening SSH publicly
for both masters and nodes, and opening port 443 on the master for the HTTPS
API endpoints.
includes intercommunication between masters and nodes, opening SSH publicly
for both masters and nodes, and opening port 443 on the master for the HTTPS
API endpoints.
* Creates an EBS volume for the master of size `MASTER_DISK_SIZE` and type
`MASTER_DISK_TYPE`.
`MASTER_DISK_TYPE`.
* Launches a master with a fixed IP address (172.20.0.9) that is also
configured for the security group and all the necessary IAM credentials. An
instance script is used to pass vital configuration information to Salt. Note:
The hope is that over time we can reduce the amount of configuration
information that must be passed in this way.
configured for the security group and all the necessary IAM credentials. An
instance script is used to pass vital configuration information to Salt. Note:
The hope is that over time we can reduce the amount of configuration
information that must be passed in this way.
* Once the instance is up, it attaches the EBS volume and sets up a manual
routing rule for the internal network range (`MASTER_IP_RANGE`, defaults to
10.246.0.0/24).
routing rule for the internal network range (`MASTER_IP_RANGE`, defaults to
10.246.0.0/24).
* For auto-scaling, on each nodes it creates a launch configuration and group.
The name for both is <*KUBE_AWS_INSTANCE_PREFIX*>-minion-group. The default
name is kubernetes-minion-group. The auto-scaling group has a min and max size
that are both set to NUM_NODES. You can change the size of the auto-scaling
group to add or remove the total number of nodes from within the AWS API or
Console. Each nodes self-configures, meaning that they come up; run Salt with
the stored configuration; connect to the master; are assigned an internal CIDR;
and then the master configures the route-table with the assigned CIDR. The
kube-up script performs a health-check on the nodes but it's a self-check that
is not required.
The name for both is <*KUBE_AWS_INSTANCE_PREFIX*>-minion-group. The default
name is kubernetes-minion-group. The auto-scaling group has a min and max size
that are both set to NUM_NODES. You can change the size of the auto-scaling
group to add or remove the total number of nodes from within the AWS API or
Console. Each nodes self-configures, meaning that they come up; run Salt with
the stored configuration; connect to the master; are assigned an internal CIDR;
and then the master configures the route-table with the assigned CIDR. The
kube-up script performs a health-check on the nodes but it's a self-check that
is not required.
If attempting this configuration manually, I highly recommend following along
If attempting this configuration manually, it is recommend to follow along
with the kube-up script, and being sure to tag everything with a tag with name
`KubernetesCluster` and value set to a unique cluster-id. Also, passing the
`KubernetesCluster` and value set to a unique cluster-id. Also, passing the
right configuration options to Salt when not using the script is tricky: the
plan here is to simplify this by having Kubernetes take on more node
configuration, and even potentially remove Salt altogether.
### Manual infrastructure creation
While this work is not yet complete, advanced users might choose to manually
create certain AWS objects while still making use of the kube-up script (to configure
Salt, for example). These objects can currently be manually created:
create certain AWS objects while still making use of the kube-up script (to
configure Salt, for example). These objects can currently be manually created:
* Set the `AWS_S3_BUCKET` environment variable to use an existing S3 bucket.
* Set the `VPC_ID` environment variable to reuse an existing VPC.
* Set the `SUBNET_ID` environment variable to reuse an existing subnet.
* If your route table has a matching `KubernetesCluster` tag, it will
be reused.
* If your route table has a matching `KubernetesCluster` tag, it will be reused.
* If your security groups are appropriately named, they will be reused.
Currently there is no way to do the following with kube-up:
* Use an existing AWS SSH key with an arbitrary name.
* Override the IAM credentials in a sensible way
([#14226](http://issues.k8s.io/14226)).
([#14226](http://issues.k8s.io/14226)).
* Use different security group permissions.
* Configure your own auto-scaling groups.
@ -312,8 +308,6 @@ Salt and Kubernetes from the S3 bucket, and then triggering Salt to actually
install Kubernetes.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->
[![Analytics](https://kubernetes-site.appspot.com/UA-36037335-10/GitHub/docs/design/aws_under_the_hood.md?pixel)]()
<!-- END MUNGE: GENERATED_ANALYTICS -->

114
docs/design/clustering.md
View File

@ -37,60 +37,122 @@ Documentation for other releases can be found at
## Overview
The term "clustering" refers to the process of having all members of the Kubernetes cluster find and trust each other. There are multiple different ways to achieve clustering with different security and usability profiles. This document attempts to lay out the user experiences for clustering that Kubernetes aims to address.
The term "clustering" refers to the process of having all members of the
Kubernetes cluster find and trust each other. There are multiple different ways
to achieve clustering with different security and usability profiles. This
document attempts to lay out the user experiences for clustering that Kubernetes
aims to address.
Once a cluster is established, the following is true:
1. **Master -> Node** The master needs to know which nodes can take work and what their current status is wrt capacity.
1. **Location** The master knows the name and location of all of the nodes in the cluster.
* For the purposes of this doc, location and name should be enough information so that the master can open a TCP connection to the Node. Most probably we will make this either an IP address or a DNS name. It is going to be important to be consistent here (master must be able to reach kubelet on that DNS name) so that we can verify certificates appropriately.
2. **Target AuthN** A way to securely talk to the kubelet on that node. Currently we call out to the kubelet over HTTP. This should be over HTTPS and the master should know what CA to trust for that node.
3. **Caller AuthN/Z** This would be the master verifying itself (and permissions) when calling the node. Currently, this is only used to collect statistics as authorization isn't critical. This may change in the future though.
2. **Node -> Master** The nodes currently talk to the master to know which pods have been assigned to them and to publish events.
1. **Master -> Node** The master needs to know which nodes can take work and
what their current status is wrt capacity.
1. **Location** The master knows the name and location of all of the nodes in
the cluster.
* For the purposes of this doc, location and name should be enough
information so that the master can open a TCP connection to the Node. Most
probably we will make this either an IP address or a DNS name. It is going to be
important to be consistent here (master must be able to reach kubelet on that
DNS name) so that we can verify certificates appropriately.
2. **Target AuthN** A way to securely talk to the kubelet on that node.
Currently we call out to the kubelet over HTTP. This should be over HTTPS and
the master should know what CA to trust for that node.
3. **Caller AuthN/Z** This would be the master verifying itself (and
permissions) when calling the node. Currently, this is only used to collect
statistics as authorization isn't critical. This may change in the future
though.
2. **Node -> Master** The nodes currently talk to the master to know which pods
have been assigned to them and to publish events.
1. **Location** The nodes must know where the master is at.
2. **Target AuthN** Since the master is assigning work to the nodes, it is critical that they verify whom they are talking to.
3. **Caller AuthN/Z** The nodes publish events and so must be authenticated to the master. Ideally this authentication is specific to each node so that authorization can be narrowly scoped. The details of the work to run (including things like environment variables) might be considered sensitive and should be locked down also.
2. **Target AuthN** Since the master is assigning work to the nodes, it is
critical that they verify whom they are talking to.
3. **Caller AuthN/Z** The nodes publish events and so must be authenticated to
the master. Ideally this authentication is specific to each node so that
authorization can be narrowly scoped. The details of the work to run (including
things like environment variables) might be considered sensitive and should be
locked down also.
**Note:** While the description here refers to a singular Master, in the future we should enable multiple Masters operating in an HA mode. While the "Master" is currently the combination of the API Server, Scheduler and Controller Manager, we will restrict ourselves to thinking about the main API and policy engine -- the API Server.
**Note:** While the description here refers to a singular Master, in the future
we should enable multiple Masters operating in an HA mode. While the "Master" is
currently the combination of the API Server, Scheduler and Controller Manager,
we will restrict ourselves to thinking about the main API and policy engine --
the API Server.
## Current Implementation
A central authority (generally the master) is responsible for determining the set of machines which are members of the cluster. Calls to create and remove worker nodes in the cluster are restricted to this single authority, and any other requests to add or remove worker nodes are rejected. (1.i).
A central authority (generally the master) is responsible for determining the
set of machines which are members of the cluster. Calls to create and remove
worker nodes in the cluster are restricted to this single authority, and any
other requests to add or remove worker nodes are rejected. (1.i.)
Communication from the master to nodes is currently over HTTP and is not secured or authenticated in any way. (1.ii, 1.iii).
Communication from the master to nodes is currently over HTTP and is not secured
or authenticated in any way. (1.ii, 1.iii.)
The location of the master is communicated out of band to the nodes. For GCE, this is done via Salt. Other cluster instructions/scripts use other methods. (2.i)
The location of the master is communicated out of band to the nodes. For GCE,
this is done via Salt. Other cluster instructions/scripts use other methods.
(2.i.)
Currently most communication from the node to the master is over HTTP. When it is done over HTTPS there is currently no verification of the cert of the master (2.ii).
Currently most communication from the node to the master is over HTTP. When it
is done over HTTPS there is currently no verification of the cert of the master
(2.ii.)
Currently, the node/kubelet is authenticated to the master via a token shared across all nodes. This token is distributed out of band (using Salt for GCE) and is optional. If it is not present then the kubelet is unable to publish events to the master. (2.iii)
Currently, the node/kubelet is authenticated to the master via a token shared
across all nodes. This token is distributed out of band (using Salt for GCE) and
is optional. If it is not present then the kubelet is unable to publish events
to the master. (2.iii.)
Our current mix of out of band communication doesn't meet all of our needs from a security point of view and is difficult to set up and configure.
Our current mix of out of band communication doesn't meet all of our needs from
a security point of view and is difficult to set up and configure.
## Proposed Solution
The proposed solution will provide a range of options for setting up and maintaining a secure Kubernetes cluster. We want to both allow for centrally controlled systems (leveraging pre-existing trust and configuration systems) or more ad-hoc automagic systems that are incredibly easy to set up.
The proposed solution will provide a range of options for setting up and
maintaining a secure Kubernetes cluster. We want to both allow for centrally
controlled systems (leveraging pre-existing trust and configuration systems) or
more ad-hoc automagic systems that are incredibly easy to set up.
The building blocks of an easier solution:
* **Move to TLS** We will move to using TLS for all intra-cluster communication. We will explicitly identify the trust chain (the set of trusted CAs) as opposed to trusting the system CAs. We will also use client certificates for all AuthN.
* [optional] **API driven CA** Optionally, we will run a CA in the master that will mint certificates for the nodes/kubelets. There will be pluggable policies that will automatically approve certificate requests here as appropriate.
* **CA approval policy** This is a pluggable policy object that can automatically approve CA signing requests. Stock policies will include `always-reject`, `queue` and `insecure-always-approve`. With `queue` there would be an API for evaluating and accepting/rejecting requests. Cloud providers could implement a policy here that verifies other out of band information and automatically approves/rejects based on other external factors.
* **Scoped Kubelet Accounts** These accounts are per-node and (optionally) give a node permission to register itself.
* To start with, we'd have the kubelets generate a cert/account in the form of `kubelet:<host>`. To start we would then hard code policy such that we give that particular account appropriate permissions. Over time, we can make the policy engine more generic.
* [optional] **Bootstrap API endpoint** This is a helper service hosted outside of the Kubernetes cluster that helps with initial discovery of the master.
* **Move to TLS** We will move to using TLS for all intra-cluster communication.
We will explicitly identify the trust chain (the set of trusted CAs) as opposed
to trusting the system CAs. We will also use client certificates for all AuthN.
* [optional] **API driven CA** Optionally, we will run a CA in the master that
will mint certificates for the nodes/kubelets. There will be pluggable policies
that will automatically approve certificate requests here as appropriate.
* **CA approval policy** This is a pluggable policy object that can
automatically approve CA signing requests. Stock policies will include
`always-reject`, `queue` and `insecure-always-approve`. With `queue` there would
be an API for evaluating and accepting/rejecting requests. Cloud providers could
implement a policy here that verifies other out of band information and
automatically approves/rejects based on other external factors.
* **Scoped Kubelet Accounts** These accounts are per-node and (optionally) give
a node permission to register itself.
* To start with, we'd have the kubelets generate a cert/account in the form of
`kubelet:<host>`. To start we would then hard code policy such that we give that
particular account appropriate permissions. Over time, we can make the policy
engine more generic.
* [optional] **Bootstrap API endpoint** This is a helper service hosted outside
of the Kubernetes cluster that helps with initial discovery of the master.
### Static Clustering
In this sequence diagram there is out of band admin entity that is creating all certificates and distributing them. It is also making sure that the kubelets know where to find the master. This provides for a lot of control but is more difficult to set up as lots of information must be communicated outside of Kubernetes.
In this sequence diagram there is out of band admin entity that is creating all
certificates and distributing them. It is also making sure that the kubelets
know where to find the master. This provides for a lot of control but is more
difficult to set up as lots of information must be communicated outside of
Kubernetes.
![Static Sequence Diagram](clustering/static.png)
### Dynamic Clustering
This diagram dynamic clustering using the bootstrap API endpoint. That API endpoint is used to both find the location of the master and communicate the root CA for the master.
This diagram shows dynamic clustering using the bootstrap API endpoint. This
endpoint is used to both find the location of the master and communicate the
root CA for the master.
This flow has the admin manually approving the kubelet signing requests. This is the `queue` policy defined above.This manual intervention could be replaced by code that can verify the signing requests via other means.
This flow has the admin manually approving the kubelet signing requests. This is
the `queue` policy defined above. This manual intervention could be replaced by
code that can verify the signing requests via other means.
![Dynamic Sequence Diagram](clustering/dynamic.png)

17
docs/design/clustering/README.md
View File

@ -33,7 +33,8 @@ Documentation for other releases can be found at
<!-- END MUNGE: UNVERSIONED_WARNING -->
This directory contains diagrams for the clustering design doc.
This depends on the `seqdiag` [utility](http://blockdiag.com/en/seqdiag/index.html). Assuming you have a non-borked python install, this should be installable with
This depends on the `seqdiag` [utility](http://blockdiag.com/en/seqdiag/index.html).
Assuming you have a non-borked python install, this should be installable with:
```sh
pip install seqdiag
@ -43,7 +44,8 @@ Just call `make` to regenerate the diagrams.
## Building with Docker
If you are on a Mac or your pip install is messed up, you can easily build with docker.
If you are on a Mac or your pip install is messed up, you can easily build with
docker:
```sh
make docker
@ -51,13 +53,18 @@ make docker
The first run will be slow but things should be fast after that.
To clean up the docker containers that are created (and other cruft that is left around) you can run `make docker-clean`.
To clean up the docker containers that are created (and other cruft that is left
around) you can run `make docker-clean`.
If you are using boot2docker and get warnings about clock skew (or if things aren't building for some reason) then you can fix that up with `make fix-clock-skew`.
If you are using boot2docker and get warnings about clock skew (or if things
aren't building for some reason) then you can fix that up with
`make fix-clock-skew`.
## Automatically rebuild on file changes
If you have the fswatch utility installed, you can have it monitor the file system and automatically rebuild when files have changed. Just do a `make watch`.
If you have the fswatch utility installed, you can have it monitor the file
system and automatically rebuild when files have changed. Just do a
`make watch`.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

53
docs/design/command_execution_port_forwarding.md
View File

@ -36,14 +36,13 @@ Documentation for other releases can be found at
## Abstract
This describes an approach for providing support for:
- executing commands in containers, with stdin/stdout/stderr streams attached
- port forwarding to containers
This document describes how to use Kubernetes to execute commands in containers,
with stdin/stdout/stderr streams attached and how to implement port forwarding
to the containers.
## Background
There are several related issues/PRs:
See the following related issues/PRs:
- [Support attach](http://issue.k8s.io/1521)
- [Real container ssh](http://issue.k8s.io/1513)
@ -77,34 +76,39 @@ won't be able to work with this mechanism, unless adapters can be written.
## Constraints and Assumptions
- SSH support is not currently in scope
- CGroup confinement is ultimately desired, but implementing that support is not currently in scope
- SELinux confinement is ultimately desired, but implementing that support is not currently in scope
- SSH support is not currently in scope.
- CGroup confinement is ultimately desired, but implementing that support is not
currently in scope.
- SELinux confinement is ultimately desired, but implementing that support is
not currently in scope.
## Use Cases
- As a user of a Kubernetes cluster, I want to run arbitrary commands in a container, attaching my local stdin/stdout/stderr to the container
- As a user of a Kubernetes cluster, I want to be able to connect to local ports on my computer and have them forwarded to ports in the container
- A user of a Kubernetes cluster wants to run arbitrary commands in a
container with local stdin/stdout/stderr attached to the container.
- A user of a Kubernetes cluster wants to connect to local ports on his computer
and have them forwarded to ports in a container.
## Process Flow
### Remote Command Execution Flow
1. The client connects to the Kubernetes Master to initiate a remote command execution
request
2. The Master proxies the request to the Kubelet where the container lives
3. The Kubelet executes nsenter + the requested command and streams stdin/stdout/stderr back and forth between the client and the container
1. The client connects to the Kubernetes Master to initiate a remote command
execution request.
2. The Master proxies the request to the Kubelet where the container lives.
3. The Kubelet executes nsenter + the requested command and streams
stdin/stdout/stderr back and forth between the client and the container.
### Port Forwarding Flow
1. The client connects to the Kubernetes Master to initiate a remote command execution
request
2. The Master proxies the request to the Kubelet where the container lives
3. The client listens on each specified local port, awaiting local connections
4. The client connects to one of the local listening ports
4. The client notifies the Kubelet of the new connection
5. The Kubelet executes nsenter + socat and streams data back and forth between the client and the port in the container
1. The client connects to the Kubernetes Master to initiate a remote command
execution request.
2. The Master proxies the request to the Kubelet where the container lives.
3. The client listens on each specified local port, awaiting local connections.
4. The client connects to one of the local listening ports.
4. The client notifies the Kubelet of the new connection.
5. The Kubelet executes nsenter + socat and streams data back and forth between
the client and the port in the container.
## Design Considerations
@ -177,7 +181,10 @@ functionality. We need to make sure that users are not allowed to execute
remote commands or do port forwarding to containers they aren't allowed to
access.
Additional work is required to ensure that multiple command execution or port forwarding connections from different clients are not able to see each other's data. This can most likely be achieved via SELinux labeling and unique process contexts.
Additional work is required to ensure that multiple command execution or port
forwarding connections from different clients are not able to see each other's
data. This can most likely be achieved via SELinux labeling and unique process
contexts.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

122
docs/design/configmap.md
View File

@ -36,8 +36,8 @@ Documentation for other releases can be found at
## Abstract
The `ConfigMap` API resource stores data used for the configuration of applications deployed on
Kubernetes.
The `ConfigMap` API resource stores data used for the configuration of
applications deployed on Kubernetes.
The main focus of this resource is to:
@ -47,71 +47,74 @@ The main focus of this resource is to:
## Motivation
A `Secret`-like API resource is needed to store configuration data that pods can consume.
A `Secret`-like API resource is needed to store configuration data that pods can
consume.
Goals of this design:
1. Describe a `ConfigMap` API resource
2. Describe the semantics of consuming `ConfigMap` as environment variables
3. Describe the semantics of consuming `ConfigMap` as files in a volume
1. Describe a `ConfigMap` API resource.
2. Describe the semantics of consuming `ConfigMap` as environment variables.
3. Describe the semantics of consuming `ConfigMap` as files in a volume.
## Use Cases
1. As a user, I want to be able to consume configuration data as environment variables
2. As a user, I want to be able to consume configuration data as files in a volume
3. As a user, I want my view of configuration data in files to be eventually consistent with changes
to the data
1. As a user, I want to be able to consume configuration data as environment
variables.
2. As a user, I want to be able to consume configuration data as files in a
volume.
3. As a user, I want my view of configuration data in files to be eventually
consistent with changes to the data.
### Consuming `ConfigMap` as Environment Variables
Many programs read their configuration from environment variables. `ConfigMap` should be possible
to consume in environment variables. The rough series of events for consuming `ConfigMap` this way
is:
A series of events for consuming `ConfigMap` as environment variables:
1. A `ConfigMap` object is created
2. A pod that consumes the configuration data via environment variables is created
3. The pod is scheduled onto a node
4. The kubelet retrieves the `ConfigMap` resource(s) referenced by the pod and starts the container
processes with the appropriate data in environment variables
1. Create a `ConfigMap` object.
2. Create a pod to consume the configuration data via environment variables.
3. The pod is scheduled onto a node.
4. The Kubelet retrieves the `ConfigMap` resource(s) referenced by the pod and
starts the container processes with the appropriate configuration data from
environment variables.
### Consuming `ConfigMap` in Volumes
Many programs read their configuration from configuration files. `ConfigMap` should be possible
to consume in a volume. The rough series of events for consuming `ConfigMap` this way
is:
A series of events for consuming `ConfigMap` as configuration files in a volume:
1. A `ConfigMap` object is created
2. A new pod using the `ConfigMap` via the volume plugin is created
3. The pod is scheduled onto a node
4. The Kubelet creates an instance of the volume plugin and calls its `Setup()` method
5. The volume plugin retrieves the `ConfigMap` resource(s) referenced by the pod and projects
the appropriate data into the volume
1. Create a `ConfigMap` object.
2. Create a new pod using the `ConfigMap` via a volume plugin.
3. The pod is scheduled onto a node.
4. The Kubelet creates an instance of the volume plugin and calls its `Setup()`
method.
5. The volume plugin retrieves the `ConfigMap` resource(s) referenced by the pod
and projects the appropriate configuration data into the volume.
### Consuming `ConfigMap` Updates
Any long-running system has configuration that is mutated over time. Changes made to configuration
data must be made visible to pods consuming data in volumes so that they can respond to those
changes.
Any long-running system has configuration that is mutated over time. Changes
made to configuration data must be made visible to pods consuming data in
volumes so that they can respond to those changes.
The `resourceVersion` of the `ConfigMap` object will be updated by the API server every time the
object is modified. After an update, modifications will be made visible to the consumer container:
The `resourceVersion` of the `ConfigMap` object will be updated by the API
server every time the object is modified. After an update, modifications will be
made visible to the consumer container:
1. A `ConfigMap` object is created
2. A new pod using the `ConfigMap` via the volume plugin is created
3. The pod is scheduled onto a node
4. During the sync loop, the Kubelet creates an instance of the volume plugin and calls its
`Setup()` method
5. The volume plugin retrieves the `ConfigMap` resource(s) referenced by the pod and projects
the appropriate data into the volume
6. The `ConfigMap` referenced by the pod is updated
7. During the next iteration of the `syncLoop`, the Kubelet creates an instance of the volume plugin
and calls its `Setup()` method
8. The volume plugin projects the updated data into the volume atomically
1. Create a `ConfigMap` object.
2. Create a new pod using the `ConfigMap` via the volume plugin.
3. The pod is scheduled onto a node.
4. During the sync loop, the Kubelet creates an instance of the volume plugin
and calls its `Setup()` method.
5. The volume plugin retrieves the `ConfigMap` resource(s) referenced by the pod
and projects the appropriate data into the volume.
6. The `ConfigMap` referenced by the pod is updated.
7. During the next iteration of the `syncLoop`, the Kubelet creates an instance
of the volume plugin and calls its `Setup()` method.
8. The volume plugin projects the updated data into the volume atomically.
It is the consuming pod's responsibility to make use of the updated data once it is made visible.
It is the consuming pod's responsibility to make use of the updated data once it
is made visible.
Because environment variables cannot be updated without restarting a container, configuration data
consumed in environment variables will not be updated.
Because environment variables cannot be updated without restarting a container,
configuration data consumed in environment variables will not be updated.
### Advantages
@ -133,8 +136,8 @@ type ConfigMap struct {
TypeMeta `json:",inline"`
ObjectMeta `json:"metadata,omitempty"`
// Data contains the configuration data. Each key must be a valid DNS_SUBDOMAIN or leading
// dot followed by valid DNS_SUBDOMAIN.
// Data contains the configuration data. Each key must be a valid
// DNS_SUBDOMAIN or leading dot followed by valid DNS_SUBDOMAIN.
Data map[string]string `json:"data,omitempty"`
}
@ -146,7 +149,8 @@ type ConfigMapList struct {
}
```
A `Registry` implementation for `ConfigMap` will be added to `pkg/registry/configmap`.
A `Registry` implementation for `ConfigMap` will be added to
`pkg/registry/configmap`.
### Environment Variables
@ -174,8 +178,8 @@ type ConfigMapSelector struct {
### Volume Source
A new `ConfigMapVolumeSource` type of volume source containing the `ConfigMap` object will be
added to the `VolumeSource` struct in the API:
A new `ConfigMapVolumeSource` type of volume source containing the `ConfigMap`
object will be added to the `VolumeSource` struct in the API:
```go
package api
@ -209,13 +213,14 @@ type KeyToPath struct {
}
```
**Note:** The update logic used in the downward API volume plug-in will be extracted and re-used in
the volume plug-in for `ConfigMap`.
**Note:** The update logic used in the downward API volume plug-in will be
extracted and re-used in the volume plug-in for `ConfigMap`.
### Changes to Secret
We will update the Secret volume plugin to have a similar API to the new ConfigMap volume plugin.
The secret volume plugin will also begin updating secret content in the volume when secrets change.
We will update the Secret volume plugin to have a similar API to the new
`ConfigMap` volume plugin. The secret volume plugin will also begin updating
secret content in the volume when secrets change.
## Examples
@ -281,7 +286,8 @@ spec:
#### Consuming `ConfigMap` as Volumes
`redis-volume-config` is intended to be used as a volume containing a config file:
`redis-volume-config` is intended to be used as a volume containing a config
file:
```yaml
apiVersion: extensions/v1beta1
@ -320,8 +326,8 @@ spec:
## Future Improvements
In the future, we may add the ability to specify an init-container that can watch the volume
contents for updates and respond to changes when they occur.
In the future, we may add the ability to specify an init-container that can
watch the volume contents for updates and respond to changes when they occur.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->
[![Analytics](https://kubernetes-site.appspot.com/UA-36037335-10/GitHub/docs/design/configmap.md?pixel)]()

89
docs/design/control-plane-resilience.md
View File

@ -54,7 +54,7 @@ ideas.
* **High availability:** continuing to be available and work correctly
even if some components are down or uncontactable. This typically
involves multiple replicas of critical services, and a reliable way
to find available replicas. Note that it's possible (but not
to find available replicas. Note that it's possible (but not
desirable) to have high
availability properties (e.g. multiple replicas) in the absence of
self-healing properties (e.g. if a replica fails, nothing replaces
@ -109,11 +109,11 @@ ideas.
## Relative Priorities
1. **(Possibly manual) recovery from catastrophic failures:** having a Kubernetes cluster, and all
applications running inside it, disappear forever perhaps is the worst
possible failure mode. So it is critical that we be able to
recover the applications running inside a cluster from such
failures in some well-bounded time period.
1. **(Possibly manual) recovery from catastrophic failures:** having a
Kubernetes cluster, and all applications running inside it, disappear forever
perhaps is the worst possible failure mode. So it is critical that we be able to
recover the applications running inside a cluster from such failures in some
well-bounded time period.
1. In theory a cluster can be recovered by replaying all API calls
that have ever been executed against it, in order, but most
often that state has been lost, and/or is scattered across
@ -121,12 +121,12 @@ ideas.
probably infeasible.
1. In theory a cluster can also be recovered to some relatively
recent non-corrupt backup/snapshot of the disk(s) backing the
etcd cluster state. But we have no default consistent
etcd cluster state. But we have no default consistent
backup/snapshot, verification or restoration process. And we
don't routinely test restoration, so even if we did routinely
perform and verify backups, we have no hard evidence that we
can in practise effectively recover from catastrophic cluster
failure or data corruption by restoring from these backups. So
failure or data corruption by restoring from these backups. So
there's more work to be done here.
1. **Self-healing:** Most major cloud providers provide the ability to
easily and automatically replace failed virtual machines within a
@ -144,7 +144,6 @@ ideas.
addition](https://github.com/coreos/etcd/blob/master/Documentation/runtime-configuration.md#add-a-new-member)
or [backup and
recovery](https://github.com/coreos/etcd/blob/master/Documentation/admin_guide.md#disaster-recovery)).
1. and boot disks are either:
1. truely persistent (i.e. remote persistent disks), or
1. reconstructible (e.g. using boot-from-snapshot,
@ -157,7 +156,7 @@ ideas.
quorum members). In environments where cloud-assisted automatic
self-healing might be infeasible (e.g. on-premise bare-metal
deployments), it also gives cluster administrators more time to
respond (e.g. replace/repair failed machines) without incurring
respond (e.g. replace/repair failed machines) without incurring
system downtime.
## Design and Status (as of December 2015)
@ -174,7 +173,7 @@ ideas.
Multiple stateless, self-hosted, self-healing API servers behind a HA
load balancer, built out by the default "kube-up" automation on GCE,
AWS and basic bare metal (BBM). Note that the single-host approach of
AWS and basic bare metal (BBM). Note that the single-host approach of
hving etcd listen only on localhost to ensure that onyl API server can
connect to it will no longer work, so alternative security will be
needed in the regard (either using firewall rules, SSL certs, or
@ -189,13 +188,13 @@ design doc.
<td>
No scripted self-healing or HA on GCE, AWS or basic bare metal
currently exists in the OSS distro. To be clear, "no self healing"
currently exists in the OSS distro. To be clear, "no self healing"
means that even if multiple e.g. API servers are provisioned for HA
purposes, if they fail, nothing replaces them, so eventually the
system will fail. Self-healing and HA can be set up
system will fail. Self-healing and HA can be set up
manually by following documented instructions, but this is not
currently an automated process, and it is not tested as part of
continuous integration. So it's probably safest to assume that it
continuous integration. So it's probably safest to assume that it
doesn't actually work in practise.
</td>
@ -205,8 +204,8 @@ doesn't actually work in practise.
<td>
Multiple self-hosted, self healing warm standby stateless controller
managers and schedulers with leader election and automatic failover of API server
clients, automatically installed by default "kube-up" automation.
managers and schedulers with leader election and automatic failover of API
server clients, automatically installed by default "kube-up" automation.
</td>
<td>As above.</td>
@ -218,47 +217,49 @@ clients, automatically installed by default "kube-up" automation.
Multiple (3-5) etcd quorum members behind a load balancer with session
affinity (to prevent clients from being bounced from one to another).
Regarding self-healing, if a node running etcd goes down, it is always necessary to do three
things:
Regarding self-healing, if a node running etcd goes down, it is always necessary
to do three things:
<ol>
<li>allocate a new node (not necessary if running etcd as a pod, in
which case specific measures are required to prevent user pods from
interfering with system pods, for example using node selectors as
described in <A HREF=")
<li>start an etcd replica on that new node,
described in <A HREF="),
<li>start an etcd replica on that new node, and
<li>have the new replica recover the etcd state.
</ol>
In the case of local disk (which fails in concert with the machine), the etcd
state must be recovered from the other replicas. This is called <A HREF="https://github.com/coreos/etcd/blob/master/Documentation/runtime-configuration.md#add-a-new-member">dynamic member
addition</A>.
In the case of remote persistent disk, the etcd state can be recovered
by attaching the remote persistent disk to the replacement node, thus
the state is recoverable even if all other replicas are down.
state must be recovered from the other replicas. This is called
<A HREF="https://github.com/coreos/etcd/blob/master/Documentation/runtime-configuration.md#add-a-new-member">
dynamic member addition</A>.
In the case of remote persistent disk, the etcd state can be recovered by
attaching the remote persistent disk to the replacement node, thus the state is
recoverable even if all other replicas are down.
There are also significant performance differences between local disks and remote
persistent disks. For example, the <A HREF="https://cloud.google.com/compute/docs/disks/#comparison_of_disk_types">sustained throughput
local disks in GCE is approximatley 20x that of remote disks</A>.
persistent disks. For example, the
<A HREF="https://cloud.google.com/compute/docs/disks/#comparison_of_disk_types">
sustained throughput local disks in GCE is approximatley 20x that of remote
disks</A>.
Hence we suggest that self-healing be provided by remotely mounted persistent disks in
non-performance critical, single-zone cloud deployments. For
performance critical installations, faster local SSD's should be used,
in which case remounting on node failure is not an option, so
<A HREF="https://github.com/coreos/etcd/blob/master/Documentation/runtime-configuration.md ">etcd runtime configuration</A>
should be used to replace the failed machine. Similarly, for
cross-zone self-healing, cloud persistent disks are zonal, so
automatic
<A HREF="https://github.com/coreos/etcd/blob/master/Documentation/runtime-configuration.md">runtime configuration</A>
is required. Similarly, basic bare metal deployments cannot generally
rely on
remote persistent disks, so the same approach applies there.
Hence we suggest that self-healing be provided by remotely mounted persistent
disks in non-performance critical, single-zone cloud deployments. For
performance critical installations, faster local SSD's should be used, in which
case remounting on node failure is not an option, so
<A HREF="https://github.com/coreos/etcd/blob/master/Documentation/runtime-configuration.md ">
etcd runtime configuration</A> should be used to replace the failed machine.
Similarly, for cross-zone self-healing, cloud persistent disks are zonal, so
automatic <A HREF="https://github.com/coreos/etcd/blob/master/Documentation/runtime-configuration.md">
runtime configuration</A> is required. Similarly, basic bare metal deployments
cannot generally rely on remote persistent disks, so the same approach applies
there.
</td>
<td>
<A HREF="http://kubernetes.io/v1.1/docs/admin/high-availability.html">
Somewhat vague instructions exist</A>
on how to set some of this up manually in a self-hosted
configuration. But automatic bootstrapping and self-healing is not
described (and is not implemented for the non-PD cases). This all
still needs to be automated and continuously tested.
Somewhat vague instructions exist</A> on how to set some of this up manually in
a self-hosted configuration. But automatic bootstrapping and self-healing is not
described (and is not implemented for the non-PD cases). This all still needs to
be automated and continuously tested.
</td>
</tr>
</table>

148
docs/design/daemon.md
View File

@ -38,40 +38,68 @@ Documentation for other releases can be found at
**Status**: Implemented.
This document presents the design of the Kubernetes DaemonSet, describes use cases, and gives an overview of the code.
This document presents the design of the Kubernetes DaemonSet, describes use
cases, and gives an overview of the code.
## Motivation
Many users have requested for a way to run a daemon on every node in a Kubernetes cluster, or on a certain set of nodes in a cluster. This is essential for use cases such as building a sharded datastore, or running a logger on every node. In comes the DaemonSet, a way to conveniently create and manage daemon-like workloads in Kubernetes.
Many users have requested for a way to run a daemon on every node in a
Kubernetes cluster, or on a certain set of nodes in a cluster. This is essential
for use cases such as building a sharded datastore, or running a logger on every
node. In comes the DaemonSet, a way to conveniently create and manage
daemon-like workloads in Kubernetes.
## Use Cases
The DaemonSet can be used for user-specified system services, cluster-level applications with strong node ties, and Kubernetes node services. Below are example use cases in each category.
The DaemonSet can be used for user-specified system services, cluster-level
applications with strong node ties, and Kubernetes node services. Below are
example use cases in each category.
### User-Specified System Services:
Logging: Some users want a way to collect statistics about nodes in a cluster and send those logs to an external database. For example, system administrators might want to know if their machines are performing as expected, if they need to add more machines to the cluster, or if they should switch cloud providers. The DaemonSet can be used to run a data collection service (for example fluentd) on every node and send the data to a service like ElasticSearch for analysis.
Logging: Some users want a way to collect statistics about nodes in a cluster
and send those logs to an external database. For example, system administrators
might want to know if their machines are performing as expected, if they need to
add more machines to the cluster, or if they should switch cloud providers. The
DaemonSet can be used to run a data collection service (for example fluentd) on
every node and send the data to a service like ElasticSearch for analysis.
### Cluster-Level Applications
Datastore: Users might want to implement a sharded datastore in their cluster. A few nodes in the cluster, labeled ‘app=datastore’, might be responsible for storing data shards, and pods running on these nodes might serve data. This architecture requires a way to bind pods to specific nodes, so it cannot be achieved using a Replication Controller. A DaemonSet is a convenient way to implement such a datastore.
Datastore: Users might want to implement a sharded datastore in their cluster. A
few nodes in the cluster, labeled ‘app=datastore’, might be responsible for
storing data shards, and pods running on these nodes might serve data. This
architecture requires a way to bind pods to specific nodes, so it cannot be
achieved using a Replication Controller. A DaemonSet is a convenient way to
implement such a datastore.
For other uses, see the related [feature request](https://issues.k8s.io/1518)
## Functionality
The DaemonSet supports standard API features:
- create
- create
- The spec for DaemonSets has a pod template field.
- Using the pod’s nodeSelector field, DaemonSets can be restricted to operate over nodes that have a certain label. For example, suppose that in a cluster some nodes are labeled ‘app=database’. You can use a DaemonSet to launch a datastore pod on exactly those nodes labeled ‘app=database’.
- Using the pod's nodeName field, DaemonSets can be restricted to operate on a specified node.
- The PodTemplateSpec used by the DaemonSet is the same as the PodTemplateSpec used by the Replication Controller.
- The initial implementation will not guarantee that DaemonSet pods are created on nodes before other pods.
- The initial implementation of DaemonSet does not guarantee that DaemonSet pods show up on nodes (for example because of resource limitations of the node), but makes a best effort to launch DaemonSet pods (like Replication Controllers do with pods). Subsequent revisions might ensure that DaemonSet pods show up on nodes, preempting other pods if necessary.
- The DaemonSet controller adds an annotation "kubernetes.io/created-by: \<json API object reference\>"
- Using the pod’s nodeSelector field, DaemonSets can be restricted to operate
over nodes that have a certain label. For example, suppose that in a cluster
some nodes are labeled ‘app=database’. You can use a DaemonSet to launch a
datastore pod on exactly those nodes labeled ‘app=database’.
- Using the pod's nodeName field, DaemonSets can be restricted to operate on a
specified node.
- The PodTemplateSpec used by the DaemonSet is the same as the PodTemplateSpec
used by the Replication Controller.
- The initial implementation will not guarantee that DaemonSet pods are
created on nodes before other pods.
- The initial implementation of DaemonSet does not guarantee that DaemonSet
pods show up on nodes (for example because of resource limitations of the node),
but makes a best effort to launch DaemonSet pods (like Replication Controllers
do with pods). Subsequent revisions might ensure that DaemonSet pods show up on
nodes, preempting other pods if necessary.
- The DaemonSet controller adds an annotation:
```"kubernetes.io/created-by: \<json API object reference\>"```
- YAML example:
```YAML
```YAML
apiVersion: extensions/v1beta1
kind: DaemonSet
metadata:
@ -94,42 +122,83 @@ The DaemonSet supports standard API features:
name: main
```
- commands that get info
- commands that get info:
- get (e.g. kubectl get daemonsets)
- describe
- Modifiers
- delete (if --cascade=true, then first the client turns down all the pods controlled by the DaemonSet (by setting the nodeSelector to a uuid pair that is unlikely to be set on any node); then it deletes the DaemonSet; then it deletes the pods)
- Modifiers:
- delete (if --cascade=true, then first the client turns down all the pods
controlled by the DaemonSet (by setting the nodeSelector to a uuid pair that is
unlikely to be set on any node); then it deletes the DaemonSet; then it deletes
the pods)
- label
- annotate
- update operations like patch and replace (only allowed to selector and to nodeSelector and nodeName of pod template)
- DaemonSets have labels, so you could, for example, list all DaemonSets with certain labels (the same way you would for a Replication Controller).
- In general, for all the supported features like get, describe, update, etc, the DaemonSet works in a similar way to the Replication Controller. However, note that the DaemonSet and the Replication Controller are different constructs.
- annotate
- update operations like patch and replace (only allowed to selector and to
nodeSelector and nodeName of pod template)
- DaemonSets have labels, so you could, for example, list all DaemonSets
with certain labels (the same way you would for a Replication Controller).
In general, for all the supported features like get, describe, update, etc,
the DaemonSet works in a similar way to the Replication Controller. However,
note that the DaemonSet and the Replication Controller are different constructs.
### Persisting Pods
- Ordinary liveness probes specified in the pod template work to keep pods created by a DaemonSet running.
- If a daemon pod is killed or stopped, the DaemonSet will create a new replica of the daemon pod on the node.
- Ordinary liveness probes specified in the pod template work to keep pods
created by a DaemonSet running.
- If a daemon pod is killed or stopped, the DaemonSet will create a new
replica of the daemon pod on the node.
### Cluster Mutations
- When a new node is added to the cluster, the DaemonSet controller starts daemon pods on the node for DaemonSets whose pod template nodeSelectors match the node’s labels.
- Suppose the user launches a DaemonSet that runs a logging daemon on all nodes labeled “logger=fluentd”. If the user then adds the “logger=fluentd” label to a node (that did not initially have the label), the logging daemon will launch on the node. Additionally, if a user removes the label from a node, the logging daemon on that node will be killed.
- When a new node is added to the cluster, the DaemonSet controller starts
daemon pods on the node for DaemonSets whose pod template nodeSelectors match
the node’s labels.
- Suppose the user launches a DaemonSet that runs a logging daemon on all
nodes labeled “logger=fluentd”. If the user then adds the “logger=fluentd” label
to a node (that did not initially have the label), the logging daemon will
launch on the node. Additionally, if a user removes the label from a node, the
logging daemon on that node will be killed.
## Alternatives Considered
We considered several alternatives, that were deemed inferior to the approach of creating a new DaemonSet abstraction.
We considered several alternatives, that were deemed inferior to the approach of
creating a new DaemonSet abstraction.
One alternative is to include the daemon in the machine image. In this case it would run outside of Kubernetes proper, and thus not be monitored, health checked, usable as a service endpoint, easily upgradable, etc.
One alternative is to include the daemon in the machine image. In this case it
would run outside of Kubernetes proper, and thus not be monitored, health
checked, usable as a service endpoint, easily upgradable, etc.
A related alternative is to package daemons as static pods. This would address most of the problems described above, but they would still not be easily upgradable, and more generally could not be managed through the API server interface.
A related alternative is to package daemons as static pods. This would address
most of the problems described above, but they would still not be easily
upgradable, and more generally could not be managed through the API server
interface.
A third alternative is to generalize the Replication Controller. We would do something like: if you set the `replicas` field of the ReplicationConrollerSpec to -1, then it means "run exactly one replica on every node matching the nodeSelector in the pod template." The ReplicationController would pretend `replicas` had been set to some large number -- larger than the largest number of nodes ever expected in the cluster -- and would use some anti-affinity mechanism to ensure that no more than one Pod from the ReplicationController runs on any given node. There are two downsides to this approach. First, there would always be a large number of Pending pods in the scheduler (these will be scheduled onto new machines when they are added to the cluster). The second downside is more philosophical: DaemonSet and the Replication Controller are very different concepts. We believe that having small, targeted controllers for distinct purposes makes Kubernetes easier to understand and use, compared to having larger multi-functional controllers (see ["Convert ReplicationController to a plugin"](http://issues.k8s.io/3058) for some discussion of this topic).
A third alternative is to generalize the Replication Controller. We would do
something like: if you set the `replicas` field of the ReplicationConrollerSpec
to -1, then it means "run exactly one replica on every node matching the
nodeSelector in the pod template." The ReplicationController would pretend
`replicas` had been set to some large number -- larger than the largest number
of nodes ever expected in the cluster -- and would use some anti-affinity
mechanism to ensure that no more than one Pod from the ReplicationController
runs on any given node. There are two downsides to this approach. First,
there would always be a large number of Pending pods in the scheduler (these
will be scheduled onto new machines when they are added to the cluster). The
second downside is more philosophical: DaemonSet and the Replication Controller
are very different concepts. We believe that having small, targeted controllers
for distinct purposes makes Kubernetes easier to understand and use, compared to
having larger multi-functional controllers (see
["Convert ReplicationController to a plugin"](http://issues.k8s.io/3058) for
some discussion of this topic).
## Design
#### Client
- Add support for DaemonSet commands to kubectl and the client. Client code was added to client/unversioned. The main files in Kubectl that were modified are kubectl/describe.go and kubectl/stop.go, since for other calls like Get, Create, and Update, the client simply forwards the request to the backend via the REST API.
- Add support for DaemonSet commands to kubectl and the client. Client code was
added to client/unversioned. The main files in Kubectl that were modified are
kubectl/describe.go and kubectl/stop.go, since for other calls like Get, Create,
and Update, the client simply forwards the request to the backend via the REST
API.
#### Apiserver
@ -137,18 +206,29 @@ A third alternative is to generalize the Replication Controller. We would do som
- REST API calls are handled in registry/daemon
- In particular, the api server will add the object to etcd
- DaemonManager listens for updates to etcd (using Framework.informer)
- API objects for DaemonSet were created in expapi/v1/types.go and expapi/v1/register.go
- API objects for DaemonSet were created in expapi/v1/types.go and
expapi/v1/register.go
- Validation code is in expapi/validation
#### Daemon Manager
- Creates new DaemonSets when requested. Launches the corresponding daemon pod on all nodes with labels matching the new DaemonSet’s selector.
- Listens for addition of new nodes to the cluster, by setting up a framework.NewInformer that watches for the creation of Node API objects. When a new node is added, the daemon manager will loop through each DaemonSet. If the label of the node matches the selector of the DaemonSet, then the daemon manager will create the corresponding daemon pod in the new node.
- The daemon manager creates a pod on a node by sending a command to the API server, requesting for a pod to be bound to the node (the node will be specified via its hostname)
- Creates new DaemonSets when requested. Launches the corresponding daemon pod
on all nodes with labels matching the new DaemonSet’s selector.
- Listens for addition of new nodes to the cluster, by setting up a
framework.NewInformer that watches for the creation of Node API objects. When a
new node is added, the daemon manager will loop through each DaemonSet. If the
label of the node matches the selector of the DaemonSet, then the daemon manager
will create the corresponding daemon pod in the new node.
- The daemon manager creates a pod on a node by sending a command to the API
server, requesting for a pod to be bound to the node (the node will be specified
via its hostname.)
#### Kubelet
- Does not need to be modified, but health checking will occur for the daemon pods and revive the pods if they are killed (we set the pod restartPolicy to Always). We reject DaemonSet objects with pod templates that don’t have restartPolicy set to Always.
- Does not need to be modified, but health checking will occur for the daemon
pods and revive the pods if they are killed (we set the pod restartPolicy to
Always). We reject DaemonSet objects with pod templates that don’t have
restartPolicy set to Always.
## Open Issues

209
docs/design/enhance-pluggable-policy.md
View File

@ -34,33 +34,60 @@ Documentation for other releases can be found at
# Enhance Pluggable Policy
While trying to develop an authorization plugin for Kubernetes, we found a few places where API extensions would ease development and add power. There are a few goals:
1. Provide an authorization plugin that can evaluate a .Authorize() call based on the full content of the request to RESTStorage. This includes information like the full verb, the content of creates and updates, and the names of resources being acted upon.
1. Provide a way to ask whether a user is permitted to take an action without running in process with the API Authorizer. For instance, a proxy for exec calls could ask whether a user can run the exec they are requesting.
1. Provide a way to ask who can perform a given action on a given resource. This is useful for answering questions like, "who can create replication controllers in my namespace".
While trying to develop an authorization plugin for Kubernetes, we found a few
places where API extensions would ease development and add power. There are a
few goals:
1. Provide an authorization plugin that can evaluate a .Authorize() call based
on the full content of the request to RESTStorage. This includes information
like the full verb, the content of creates and updates, and the names of
resources being acted upon.
1. Provide a way to ask whether a user is permitted to take an action without
running in process with the API Authorizer. For instance, a proxy for exec
calls could ask whether a user can run the exec they are requesting.
1. Provide a way to ask who can perform a given action on a given resource.
This is useful for answering questions like, "who can create replication
controllers in my namespace".
This proposal adds to and extends the existing API to so that authorizers may provide the functionality described above. It does not attempt to describe how the policies themselves can be expressed, that is up the authorization plugins themselves.
This proposal adds to and extends the existing API to so that authorizers may
provide the functionality described above. It does not attempt to describe how
the policies themselves can be expressed, that is up the authorization plugins
themselves.
## Enhancements to existing Authorization interfaces
The existing Authorization interfaces are described here: [docs/admin/authorization.md](../admin/authorization.md). A couple additions will allow the development of an Authorizer that matches based on different rules than the existing implementation.
The existing Authorization interfaces are described
[here](../admin/authorization.md). A couple additions will allow the development
of an Authorizer that matches based on different rules than the existing
implementation.
### Request Attributes
The existing authorizer.Attributes only has 5 attributes (user, groups, isReadOnly, kind, and namespace). If we add more detailed verbs, content, and resource names, then Authorizer plugins will have the same level of information available to RESTStorage components in order to express more detailed policy. The replacement excerpt is below.
The existing authorizer.Attributes only has 5 attributes (user, groups,
isReadOnly, kind, and namespace). If we add more detailed verbs, content, and
resource names, then Authorizer plugins will have the same level of information
available to RESTStorage components in order to express more detailed policy.
The replacement excerpt is below.
An API request has the following attributes that can be considered for authorization:
- user - the user-string which a user was authenticated as. This is included in the Context.
- groups - the groups to which the user belongs. This is included in the Context.
- verb - string describing the requesting action. Today we have: get, list, watch, create, update, and delete. The old `readOnly` behavior is equivalent to allowing get, list, watch.
- namespace - the namespace of the object being access, or the empty string if the endpoint does not support namespaced objects. This is included in the Context.
An API request has the following attributes that can be considered for
authorization:
- user - the user-string which a user was authenticated as. This is included
in the Context.
- groups - the groups to which the user belongs. This is included in the
Context.
- verb - string describing the requesting action. Today we have: get, list,
watch, create, update, and delete. The old `readOnly` behavior is equivalent to
allowing get, list, watch.
- namespace - the namespace of the object being access, or the empty string if
the endpoint does not support namespaced objects. This is included in the
Context.
- resourceGroup - the API group of the resource being accessed
- resourceVersion - the API version of the resource being accessed
- resource - which resource is being accessed
- applies only to the API endpoints, such as
`/api/v1beta1/pods`. For miscellaneous endpoints, like `/version`, the kind is the empty string.
- resourceName - the name of the resource during a get, update, or delete action.
- applies only to the API endpoints, such as `/api/v1beta1/pods`. For
miscellaneous endpoints, like `/version`, the kind is the empty string.
- resourceName - the name of the resource during a get, update, or delete
action.
- subresource - which subresource is being accessed
A non-API request has 2 attributes:
@ -70,7 +97,14 @@ A non-API request has 2 attributes:
### Authorizer Interface
The existing Authorizer interface is very simple, but there isn't a way to provide details about allows, denies, or failures. The extended detail is useful for UIs that want to describe why certain actions are allowed or disallowed. Not all Authorizers will want to provide that information, but for those that do, having that capability is useful. In addition, adding a `GetAllowedSubjects` method that returns back the users and groups that can perform a particular action makes it possible to answer questions like, "who can see resources in my namespace" (see [ResourceAccessReview](#ResourceAccessReview) further down).
The existing Authorizer interface is very simple, but there isn't a way to
provide details about allows, denies, or failures. The extended detail is useful
for UIs that want to describe why certain actions are allowed or disallowed. Not
all Authorizers will want to provide that information, but for those that do,
having that capability is useful. In addition, adding a `GetAllowedSubjects`
method that returns back the users and groups that can perform a particular
action makes it possible to answer questions like, "who can see resources in my
namespace" (see [ResourceAccessReview](#ResourceAccessReview) further down).
```go
// OLD
@ -81,41 +115,65 @@ type Authorizer interface {
```go
// NEW
// Authorizer provides the ability to determine if a particular user can perform a particular action
// Authorizer provides the ability to determine if a particular user can perform
// a particular action
type Authorizer interface {
// Authorize takes a Context (for namespace, user, and traceability) and Attributes to make a policy determination.
// reason is an optional return value that can describe why a policy decision was made. Reasons are useful during
// debugging when trying to figure out why a user or group has access to perform a particular action.
// Authorize takes a Context (for namespace, user, and traceability) and
// Attributes to make a policy determination.
// reason is an optional return value that can describe why a policy decision
// was made. Reasons are useful during debugging when trying to figure out
// why a user or group has access to perform a particular action.
Authorize(ctx api.Context, a Attributes) (allowed bool, reason string, evaluationError error)
}
// AuthorizerIntrospection is an optional interface that provides the ability to determine which users and groups can perform a particular action.
// This is useful for building caches of who can see what: for instance, "which namespaces can this user see". That would allow
// someone to see only the namespaces they are allowed to view instead of having to choose between listing them all or listing none.
// AuthorizerIntrospection is an optional interface that provides the ability to
// determine which users and groups can perform a particular action. This is
// useful for building caches of who can see what. For instance, "which
// namespaces can this user see". That would allow someone to see only the
// namespaces they are allowed to view instead of having to choose between
// listing them all or listing none.
type AuthorizerIntrospection interface {
// GetAllowedSubjects takes a Context (for namespace and traceability) and Attributes to determine which users and
// groups are allowed to perform the described action in the namespace. This API enables the ResourceBasedReview requests below
// GetAllowedSubjects takes a Context (for namespace and traceability) and
// Attributes to determine which users and groups are allowed to perform the
// described action in the namespace. This API enables the ResourceBasedReview
// requests below
GetAllowedSubjects(ctx api.Context, a Attributes) (users util.StringSet, groups util.StringSet, evaluationError error)
}
```
### SubjectAccessReviews
This set of APIs answers the question: can a user or group (use authenticated user if none is specified) perform a given action. Given the Authorizer interface (proposed or existing), this endpoint can be implemented generically against any Authorizer by creating the correct Attributes and making an .Authorize() call.
This set of APIs answers the question: can a user or group (use authenticated
user if none is specified) perform a given action. Given the Authorizer
interface (proposed or existing), this endpoint can be implemented generically
against any Authorizer by creating the correct Attributes and making an
.Authorize() call.
There are three different flavors:
1. `/apis/authorization.kubernetes.io/{version}/subjectAccessReviews` - this checks to see if a specified user or group can perform a given action at the cluster scope or across all namespaces.
This is a highly privileged operation. It allows a cluster-admin to inspect rights of any person across the entire cluster and against cluster level resources.
2. `/apis/authorization.kubernetes.io/{version}/personalSubjectAccessReviews` - this checks to see if the current user (including his groups) can perform a given action at any specified scope.
This is an unprivileged operation. It doesn't expose any information that a user couldn't discover simply by trying an endpoint themselves.
3. `/apis/authorization.kubernetes.io/{version}/ns/{namespace}/localSubjectAccessReviews` - this checks to see if a specified user or group can perform a given action in **this** namespace.
This is a moderately privileged operation. In a multi-tenant environment, have a namespace scoped resource makes it very easy to reason about powers granted to a namespace admin.
This allows a namespace admin (someone able to manage permissions inside of one namespaces, but not all namespaces), the power to inspect whether a given user or group
can manipulate resources in his namespace.
1. `/apis/authorization.kubernetes.io/{version}/subjectAccessReviews` - this
checks to see if a specified user or group can perform a given action at the
cluster scope or across all namespaces. This is a highly privileged operation.
It allows a cluster-admin to inspect rights of any person across the entire
cluster and against cluster level resources.
2. `/apis/authorization.kubernetes.io/{version}/personalSubjectAccessReviews` -
this checks to see if the current user (including his groups) can perform a
given action at any specified scope. This is an unprivileged operation. It
doesn't expose any information that a user couldn't discover simply by trying an
endpoint themselves.
3. `/apis/authorization.kubernetes.io/{version}/ns/{namespace}/localSubjectAccessReviews` -
this checks to see if a specified user or group can perform a given action in
**this** namespace. This is a moderately privileged operation. In a multi-tenant
environment, having a namespace scoped resource makes it very easy to reason
about powers granted to a namespace admin. This allows a namespace admin
(someone able to manage permissions inside of one namespaces, but not all
namespaces), the power to inspect whether a given user or group can manipulate
resources in his namespace.
SubjectAccessReview is runtime.Object with associated RESTStorage that only accepts creates. The caller POSTs a SubjectAccessReview to this URL and he gets a SubjectAccessReviewResponse back. Here is an example of a call and its corresponding return.
SubjectAccessReview is runtime.Object with associated RESTStorage that only
accepts creates. The caller POSTs a SubjectAccessReview to this URL and he gets
a SubjectAccessReviewResponse back. Here is an example of a call and its
corresponding return:
```
// input
@ -141,10 +199,14 @@ accessReviewResult, err := Client.SubjectAccessReviews().Create(subjectAccessRev
"apiVersion": "authorization.kubernetes.io/v1",
"allowed": true
}
PersonalSubjectAccessReview is runtime.Object with associated RESTStorage that only accepts creates. The caller POSTs a PersonalSubjectAccessReview to this URL and he gets a SubjectAccessReviewResponse back. Here is an example of a call and its corresponding return.
```
PersonalSubjectAccessReview is runtime.Object with associated RESTStorage that
only accepts creates. The caller POSTs a PersonalSubjectAccessReview to this URL
and he gets a SubjectAccessReviewResponse back. Here is an example of a call and
its corresponding return:
```
// input
{
"kind": "PersonalSubjectAccessReview",
@ -167,8 +229,12 @@ accessReviewResult, err := Client.PersonalSubjectAccessReviews().Create(subjectA
"apiVersion": "authorization.kubernetes.io/v1",
"allowed": true
}
```
LocalSubjectAccessReview is runtime.Object with associated RESTStorage that only accepts creates. The caller POSTs a LocalSubjectAccessReview to this URL and he gets a LocalSubjectAccessReviewResponse back. Here is an example of a call and its corresponding return.
LocalSubjectAccessReview is runtime.Object with associated RESTStorage that only
accepts creates. The caller POSTs a LocalSubjectAccessReview to this URL and he
gets a LocalSubjectAccessReviewResponse back. Here is an example of a call and
its corresponding return:
```
// input
@ -196,15 +262,14 @@ accessReviewResult, err := Client.LocalSubjectAccessReviews().Create(localSubjec
"namespace": "my-ns"
"allowed": true
}
```
The actual Go objects look like this:
```go
type AuthorizationAttributes struct {
// Namespace is the namespace of the action being requested. Currently, there is no distinction between no namespace and all namespaces
// Namespace is the namespace of the action being requested. Currently, there
// is no distinction between no namespace and all namespaces
Namespace string `json:"namespace" description:"namespace of the action being requested"`
// Verb is one of: get, list, watch, create, update, delete
Verb string `json:"verb" description:"one of get, list, watch, create, update, delete"`
@ -214,13 +279,15 @@ type AuthorizationAttributes struct {
ResourceVersion string `json:"resourceVersion" description:"version of the resource being requested"`
// Resource is one of the existing resource types
Resource string `json:"resource" description:"one of the existing resource types"`
// ResourceName is the name of the resource being requested for a "get" or deleted for a "delete"
// ResourceName is the name of the resource being requested for a "get" or
// deleted for a "delete"
ResourceName string `json:"resourceName" description:"name of the resource being requested for a get or delete"`
// Subresource is one of the existing subresources types
Subresource string `json:"subresource" description:"one of the existing subresources"`
}
// SubjectAccessReview is an object for requesting information about whether a user or group can perform an action
// SubjectAccessReview is an object for requesting information about whether a
// user or group can perform an action
type SubjectAccessReview struct {
kapi.TypeMeta `json:",inline"`
@ -232,7 +299,8 @@ type SubjectAccessReview struct {
Groups []string `json:"groups" description:"optional, list of groups to which the user belongs"`
}
// SubjectAccessReviewResponse describes whether or not a user or group can perform an action
// SubjectAccessReviewResponse describes whether or not a user or group can
// perform an action
type SubjectAccessReviewResponse struct {
kapi.TypeMeta
@ -242,7 +310,8 @@ type SubjectAccessReviewResponse struct {
Reason string
}
// PersonalSubjectAccessReview is an object for requesting information about whether a user or group can perform an action
// PersonalSubjectAccessReview is an object for requesting information about
// whether a user or group can perform an action
type PersonalSubjectAccessReview struct {
kapi.TypeMeta `json:",inline"`
@ -250,7 +319,8 @@ type PersonalSubjectAccessReview struct {
AuthorizationAttributes `json:"authorizationAttributes" description:"the action being tested"`
}
// PersonalSubjectAccessReviewResponse describes whether this user can perform an action
// PersonalSubjectAccessReviewResponse describes whether this user can perform
// an action
type PersonalSubjectAccessReviewResponse struct {
kapi.TypeMeta
@ -262,7 +332,8 @@ type PersonalSubjectAccessReviewResponse struct {
Reason string
}
// LocalSubjectAccessReview is an object for requesting information about whether a user or group can perform an action
// LocalSubjectAccessReview is an object for requesting information about
// whether a user or group can perform an action
type LocalSubjectAccessReview struct {
kapi.TypeMeta `json:",inline"`
@ -274,7 +345,8 @@ type LocalSubjectAccessReview struct {
Groups []string `json:"groups" description:"optional, list of groups to which the user belongs"`
}
// LocalSubjectAccessReviewResponse describes whether or not a user or group can perform an action
// LocalSubjectAccessReviewResponse describes whether or not a user or group can
// perform an action
type LocalSubjectAccessReviewResponse struct {
kapi.TypeMeta
@ -287,21 +359,33 @@ type LocalSubjectAccessReviewResponse struct {
}
```
### ResourceAccessReview
This set of APIs nswers the question: which users and groups can perform the specified verb on the specified resourceKind. Given the Authorizer interface described above, this endpoint can be implemented generically against any Authorizer by calling the .GetAllowedSubjects() function.
This set of APIs nswers the question: which users and groups can perform the
specified verb on the specified resourceKind. Given the Authorizer interface
described above, this endpoint can be implemented generically against any
Authorizer by calling the .GetAllowedSubjects() function.
There are two different flavors:
1. `/apis/authorization.kubernetes.io/{version}/resourceAccessReview` - this checks to see which users and groups can perform a given action at the cluster scope or across all namespaces.
This is a highly privileged operation. It allows a cluster-admin to inspect rights of all subjects across the entire cluster and against cluster level resources.
2. `/apis/authorization.kubernetes.io/{version}/ns/{namespace}/localResourceAccessReviews` - this checks to see which users and groups can perform a given action in **this** namespace.
This is a moderately privileged operation. In a multi-tenant environment, have a namespace scoped resource makes it very easy to reason about powers granted to a namespace admin.
This allows a namespace admin (someone able to manage permissions inside of one namespaces, but not all namespaces), the power to inspect which users and groups
can manipulate resources in his namespace.
1. `/apis/authorization.kubernetes.io/{version}/resourceAccessReview` - this
checks to see which users and groups can perform a given action at the cluster
scope or across all namespaces. This is a highly privileged operation. It allows
a cluster-admin to inspect rights of all subjects across the entire cluster and
against cluster level resources.
2. `/apis/authorization.kubernetes.io/{version}/ns/{namespace}/localResourceAccessReviews` -
this checks to see which users and groups can perform a given action in **this**
namespace. This is a moderately privileged operation. In a multi-tenant
environment, having a namespace scoped resource makes it very easy to reason
about powers granted to a namespace admin. This allows a namespace admin
(someone able to manage permissions inside of one namespaces, but not all
namespaces), the power to inspect which users and groups can manipulate
resources in his namespace.
ResourceAccessReview is a runtime.Object with associated RESTStorage that only accepts creates. The caller POSTs a ResourceAccessReview to this URL and he gets a ResourceAccessReviewResponse back. Here is an example of a call and its corresponding return.
ResourceAccessReview is a runtime.Object with associated RESTStorage that only
accepts creates. The caller POSTs a ResourceAccessReview to this URL and he gets
a ResourceAccessReviewResponse back. Here is an example of a call and its
corresponding return:
```
// input
@ -332,8 +416,8 @@ accessReviewResult, err := Client.ResourceAccessReviews().Create(resourceAccessR
The actual Go objects look like this:
```go
// ResourceAccessReview is a means to request a list of which users and groups are authorized to perform the
// action specified by spec
// ResourceAccessReview is a means to request a list of which users and groups
// are authorized to perform the action specified by spec
type ResourceAccessReview struct {
kapi.TypeMeta `json:",inline"`
@ -351,8 +435,8 @@ type ResourceAccessReviewResponse struct {
Groups []string
}
// LocalResourceAccessReview is a means to request a list of which users and groups are authorized to perform the
// action specified in a specific namespace
// LocalResourceAccessReview is a means to request a list of which users and
// groups are authorized to perform the action specified in a specific namespace
type LocalResourceAccessReview struct {
kapi.TypeMeta `json:",inline"`
@ -371,7 +455,6 @@ type LocalResourceAccessReviewResponse struct {
// Groups is the list of groups who can perform the action
Groups []string
}
```

132
docs/design/event_compression.md
View File

@ -42,40 +42,62 @@ Kubernetes components can get into a state where they generate tons of events.
The events can be categorized in one of two ways:
1. same - the event is identical to previous events except it varies only on timestamp
2. similar - the event is identical to previous events except it varies on timestamp and message
1. same - The event is identical to previous events except it varies only on
timestamp.
2. similar - The event is identical to previous events except it varies on
timestamp and message.
For example, when pulling a non-existing image, Kubelet will repeatedly generate `image_not_existing` and `container_is_waiting` events until upstream components correct the image. When this happens, the spam from the repeated events makes the entire event mechanism useless. It also appears to cause memory pressure in etcd (see [#3853](http://issue.k8s.io/3853)).
For example, when pulling a non-existing image, Kubelet will repeatedly generate
`image_not_existing` and `container_is_waiting` events until upstream components
correct the image. When this happens, the spam from the repeated events makes
the entire event mechanism useless. It also appears to cause memory pressure in
etcd (see [#3853](http://issue.k8s.io/3853)).
The goal is introduce event counting to increment same events, and event aggregation to collapse similar events.
The goal is introduce event counting to increment same events, and event
aggregation to collapse similar events.
## Proposal
Each binary that generates events (for example, `kubelet`) should keep track of previously generated events so that it can collapse recurring events into a single event instead of creating a new instance for each new event. In addition, if many similar events are
created, events should be aggregated into a single event to reduce spam.
Each binary that generates events (for example, `kubelet`) should keep track of
previously generated events so that it can collapse recurring events into a
single event instead of creating a new instance for each new event. In addition,
if many similar events are created, events should be aggregated into a single
event to reduce spam.
Event compression should be best effort (not guaranteed). Meaning, in the worst case, `n` identical (minus timestamp) events may still result in `n` event entries.
Event compression should be best effort (not guaranteed). Meaning, in the worst
case, `n` identical (minus timestamp) events may still result in `n` event
entries.
## Design
Instead of a single Timestamp, each event object [contains](http://releases.k8s.io/HEAD/pkg/api/types.go#L1111) the following fields:
Instead of a single Timestamp, each event object
[contains](http://releases.k8s.io/HEAD/pkg/api/types.go#L1111) the following
fields:
* `FirstTimestamp unversioned.Time`
* The date/time of the first occurrence of the event.
* `LastTimestamp unversioned.Time`
* The date/time of the most recent occurrence of the event.
* On first occurrence, this is equal to the FirstTimestamp.
* `Count int`
* The number of occurrences of this event between FirstTimestamp and LastTimestamp
* The number of occurrences of this event between FirstTimestamp and
LastTimestamp.
* On first occurrence, this is 1.
Each binary that generates events:
* Maintains a historical record of previously generated events:
* Implemented with ["Least Recently Used Cache"](https://github.com/golang/groupcache/blob/master/lru/lru.go) in [`pkg/client/record/events_cache.go`](../../pkg/client/record/events_cache.go).
* Implemented behind an `EventCorrelator` that manages two subcomponents: `EventAggregator` and `EventLogger`
* The `EventCorrelator` observes all incoming events and lets each subcomponent visit and modify the event in turn.
* The `EventAggregator` runs an aggregation function over each event. This function buckets each event based on an `aggregateKey`,
and identifies the event uniquely with a `localKey` in that bucket.
* The default aggregation function groups similar events that differ only by `event.Message`. It's `localKey` is `event.Message` and its aggregate key is produced by joining:
* Implemented with
["Least Recently Used Cache"](https://github.com/golang/groupcache/blob/master/lru/lru.go)
in [`pkg/client/record/events_cache.go`](../../pkg/client/record/events_cache.go).
* Implemented behind an `EventCorrelator` that manages two subcomponents:
`EventAggregator` and `EventLogger`.
* The `EventCorrelator` observes all incoming events and lets each
subcomponent visit and modify the event in turn.
* The `EventAggregator` runs an aggregation function over each event. This
function buckets each event based on an `aggregateKey` and identifies the event
uniquely with a `localKey` in that bucket.
* The default aggregation function groups similar events that differ only by
`event.Message`. Its `localKey` is `event.Message` and its aggregate key is
produced by joining:
* `event.Source.Component`
* `event.Source.Host`
* `event.InvolvedObject.Kind`
@ -84,12 +106,17 @@ Each binary that generates events:
* `event.InvolvedObject.UID`
* `event.InvolvedObject.APIVersion`
* `event.Reason`
* If the `EventAggregator` observes a similar event produced 10 times in a 10 minute window, it drops the event that was provided as
input and creates a new event that differs only on the message. The message denotes that this event is used to group similar events
that matched on reason. This aggregated `Event` is then used in the event processing sequence.
* The `EventLogger` observes the event out of `EventAggregation` and tracks the number of times it has observed that event previously
by incrementing a key in a cache associated with that matching event.
* The key in the cache is generated from the event object minus timestamps/count/transient fields, specifically the following events fields are used to construct a unique key for an event:
* If the `EventAggregator` observes a similar event produced 10 times in a 10
minute window, it drops the event that was provided as input and creates a new
event that differs only on the message. The message denotes that this event is
used to group similar events that matched on reason. This aggregated `Event` is
then used in the event processing sequence.
* The `EventLogger` observes the event out of `EventAggregation` and tracks
the number of times it has observed that event previously by incrementing a key
in a cache associated with that matching event.
* The key in the cache is generated from the event object minus
timestamps/count/transient fields, specifically the following events fields are
used to construct a unique key for an event:
* `event.Source.Component`
* `event.Source.Host`
* `event.InvolvedObject.Kind`
@ -99,24 +126,47 @@ Each binary that generates events:
* `event.InvolvedObject.APIVersion`
* `event.Reason`
* `event.Message`
* The LRU cache is capped at 4096 events for both `EventAggregator` and `EventLogger`. That means if a component (e.g. kubelet) runs for a long period of time and generates tons of unique events, the previously generated events cache will not grow unchecked in memory. Instead, after 4096 unique events are generated, the oldest events are evicted from the cache.
* When an event is generated, the previously generated events cache is checked (see [`pkg/client/unversioned/record/event.go`](http://releases.k8s.io/HEAD/pkg/client/record/event.go)).
* If the key for the new event matches the key for a previously generated event (meaning all of the above fields match between the new event and some previously generated event), then the event is considered to be a duplicate and the existing event entry is updated in etcd:
* The new PUT (update) event API is called to update the existing event entry in etcd with the new last seen timestamp and count.
* The event is also updated in the previously generated events cache with an incremented count, updated last seen timestamp, name, and new resource version (all required to issue a future event update).
* If the key for the new event does not match the key for any previously generated event (meaning none of the above fields match between the new event and any previously generated events), then the event is considered to be new/unique and a new event entry is created in etcd:
* The usual POST/create event API is called to create a new event entry in etcd.
* An entry for the event is also added to the previously generated events cache.
* The LRU cache is capped at 4096 events for both `EventAggregator` and
`EventLogger`. That means if a component (e.g. kubelet) runs for a long period
of time and generates tons of unique events, the previously generated events
cache will not grow unchecked in memory. Instead, after 4096 unique events are
generated, the oldest events are evicted from the cache.
* When an event is generated, the previously generated events cache is checked
(see [`pkg/client/unversioned/record/event.go`](http://releases.k8s.io/HEAD/pkg/client/record/event.go)).
* If the key for the new event matches the key for a previously generated
event (meaning all of the above fields match between the new event and some
previously generated event), then the event is considered to be a duplicate and
the existing event entry is updated in etcd:
* The new PUT (update) event API is called to update the existing event
entry in etcd with the new last seen timestamp and count.
* The event is also updated in the previously generated events cache with
an incremented count, updated last seen timestamp, name, and new resource
version (all required to issue a future event update).
* If the key for the new event does not match the key for any previously
generated event (meaning none of the above fields match between the new event
and any previously generated events), then the event is considered to be
new/unique and a new event entry is created in etcd:
* The usual POST/create event API is called to create a new event entry in
etcd.
* An entry for the event is also added to the previously generated events
cache.
## Issues/Risks
* Compression is not guaranteed, because each component keeps track of event history in memory
* An application restart causes event history to be cleared, meaning event history is not preserved across application restarts and compression will not occur across component restarts.
* Because an LRU cache is used to keep track of previously generated events, if too many unique events are generated, old events will be evicted from the cache, so events will only be compressed until they age out of the events cache, at which point any new instance of the event will cause a new entry to be created in etcd.
* Compression is not guaranteed, because each component keeps track of event
history in memory
* An application restart causes event history to be cleared, meaning event
history is not preserved across application restarts and compression will not
occur across component restarts.
* Because an LRU cache is used to keep track of previously generated events,
if too many unique events are generated, old events will be evicted from the
cache, so events will only be compressed until they age out of the events cache,
at which point any new instance of the event will cause a new entry to be
created in etcd.
## Example
Sample kubectl output
Sample kubectl output:
```console
FIRSTSEEN LASTSEEN COUNT NAME KIND SUBOBJECT REASON SOURCE MESSAGE
@ -133,15 +183,19 @@ Thu, 12 Feb 2015 01:13:20 +0000 Thu, 12 Feb 2015 01:13:20 +0000 1
Thu, 12 Feb 2015 01:13:20 +0000 Thu, 12 Feb 2015 01:13:20 +0000 1 kibana-logging-controller-gziey Pod scheduled {scheduler } Successfully assigned kibana-logging-controller-gziey to kubernetes-node-4.c.saad-dev-vms.internal
```
This demonstrates what would have been 20 separate entries (indicating scheduling failure) collapsed/compressed down to 5 entries.
This demonstrates what would have been 20 separate entries (indicating
scheduling failure) collapsed/compressed down to 5 entries.
## Related Pull Requests/Issues
* Issue [#4073](http://issue.k8s.io/4073): Compress duplicate events
* PR [#4157](http://issue.k8s.io/4157): Add "Update Event" to Kubernetes API
* PR [#4206](http://issue.k8s.io/4206): Modify Event struct to allow compressing multiple recurring events in to a single event
* PR [#4306](http://issue.k8s.io/4306): Compress recurring events in to a single event to optimize etcd storage
* PR [#4444](http://pr.k8s.io/4444): Switch events history to use LRU cache instead of map
* Issue [#4073](http://issue.k8s.io/4073): Compress duplicate events.
* PR [#4157](http://issue.k8s.io/4157): Add "Update Event" to Kubernetes API.
* PR [#4206](http://issue.k8s.io/4206): Modify Event struct to allow
compressing multiple recurring events in to a single event.
* PR [#4306](http://issue.k8s.io/4306): Compress recurring events in to a
single event to optimize etcd storage.
* PR [#4444](http://pr.k8s.io/4444): Switch events history to use LRU cache
instead of map.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

278
docs/design/expansion.md
View File

@ -36,13 +36,15 @@ Documentation for other releases can be found at
## Abstract
A proposal for the expansion of environment variables using a simple `$(var)` syntax.
A proposal for the expansion of environment variables using a simple `$(var)`
syntax.
## Motivation
It is extremely common for users to need to compose environment variables or pass arguments to
their commands using the values of environment variables. Kubernetes should provide a facility for
the 80% cases in order to decrease coupling and the use of workarounds.
It is extremely common for users to need to compose environment variables or
pass arguments to their commands using the values of environment variables.
Kubernetes should provide a facility for the 80% cases in order to decrease
coupling and the use of workarounds.
## Goals
@ -53,150 +55,170 @@ the 80% cases in order to decrease coupling and the use of workarounds.
## Constraints and Assumptions
* This design should describe the simplest possible syntax to accomplish the use-cases
* Expansion syntax will not support more complicated shell-like behaviors such as default values
(viz: `$(VARIABLE_NAME:"default")`), inline substitution, etc.
* This design should describe the simplest possible syntax to accomplish the
use-cases.
* Expansion syntax will not support more complicated shell-like behaviors such
as default values (viz: `$(VARIABLE_NAME:"default")`), inline substitution, etc.
## Use Cases
1. As a user, I want to compose new environment variables for a container using a substitution
syntax to reference other variables in the container's environment and service environment
variables
1. As a user, I want to substitute environment variables into a container's command
1. As a user, I want to do the above without requiring the container's image to have a shell
1. As a user, I want to be able to specify a default value for a service variable which may
not exist
1. As a user, I want to see an event associated with the pod if an expansion fails (ie, references
variable names that cannot be expanded)
1. As a user, I want to compose new environment variables for a container using
a substitution syntax to reference other variables in the container's
environment and service environment variables.
1. As a user, I want to substitute environment variables into a container's
command.
1. As a user, I want to do the above without requiring the container's image to
have a shell.
1. As a user, I want to be able to specify a default value for a service
variable which may not exist.
1. As a user, I want to see an event associated with the pod if an expansion
fails (ie, references variable names that cannot be expanded).
### Use Case: Composition of environment variables
Currently, containers are injected with docker-style environment variables for the services in
their pod's namespace. There are several variables for each service, but users routinely need
to compose URLs based on these variables because there is not a variable for the exact format
they need. Users should be able to build new environment variables with the exact format they need.
Eventually, it should also be possible to turn off the automatic injection of the docker-style
variables into pods and let the users consume the exact information they need via the downward API
and composition.
Currently, containers are injected with docker-style environment variables for
the services in their pod's namespace. There are several variables for each
service, but users routinely need to compose URLs based on these variables
because there is not a variable for the exact format they need. Users should be
able to build new environment variables with the exact format they need.
Eventually, it should also be possible to turn off the automatic injection of
the docker-style variables into pods and let the users consume the exact
information they need via the downward API and composition.
#### Expanding expanded variables
It should be possible to reference an variable which is itself the result of an expansion, if the
referenced variable is declared in the container's environment prior to the one referencing it.
Put another way -- a container's environment is expanded in order, and expanded variables are
available to subsequent expansions.
It should be possible to reference an variable which is itself the result of an
expansion, if the referenced variable is declared in the container's environment
prior to the one referencing it. Put another way -- a container's environment is
expanded in order, and expanded variables are available to subsequent
expansions.
### Use Case: Variable expansion in command
Users frequently need to pass the values of environment variables to a container's command.
Currently, Kubernetes does not perform any expansion of variables. The workaround is to invoke a
shell in the container's command and have the shell perform the substitution, or to write a wrapper
script that sets up the environment and runs the command. This has a number of drawbacks:
Users frequently need to pass the values of environment variables to a
container's command. Currently, Kubernetes does not perform any expansion of
variables. The workaround is to invoke a shell in the container's command and
have the shell perform the substitution, or to write a wrapper script that sets
up the environment and runs the command. This has a number of drawbacks:
1. Solutions that require a shell are unfriendly to images that do not contain a shell
2. Wrapper scripts make it harder to use images as base images
3. Wrapper scripts increase coupling to Kubernetes
1. Solutions that require a shell are unfriendly to images that do not contain
a shell.
2. Wrapper scripts make it harder to use images as base images.
3. Wrapper scripts increase coupling to Kubernetes.
Users should be able to do the 80% case of variable expansion in command without writing a wrapper
script or adding a shell invocation to their containers' commands.
Users should be able to do the 80% case of variable expansion in command without
writing a wrapper script or adding a shell invocation to their containers'
commands.
### Use Case: Images without shells
The current workaround for variable expansion in a container's command requires the container's
image to have a shell. This is unfriendly to images that do not contain a shell (`scratch` images,
for example). Users should be able to perform the other use-cases in this design without regard to
the content of their images.
The current workaround for variable expansion in a container's command requires
the container's image to have a shell. This is unfriendly to images that do not
contain a shell (`scratch` images, for example). Users should be able to perform
the other use-cases in this design without regard to the content of their
images.
### Use Case: See an event for incomplete expansions
It is possible that a container with incorrect variable values or command line may continue to run
for a long period of time, and that the end-user would have no visual or obvious warning of the
incorrect configuration. If the kubelet creates an event when an expansion references a variable
that cannot be expanded, it will help users quickly detect problems with expansions.
It is possible that a container with incorrect variable values or command line
may continue to run for a long period of time, and that the end-user would have
no visual or obvious warning of the incorrect configuration. If the kubelet
creates an event when an expansion references a variable that cannot be
expanded, it will help users quickly detect problems with expansions.
## Design Considerations
### What features should be supported?
In order to limit complexity, we want to provide the right amount of functionality so that the 80%
cases can be realized and nothing more. We felt that the essentials boiled down to:
In order to limit complexity, we want to provide the right amount of
functionality so that the 80% cases can be realized and nothing more. We felt
that the essentials boiled down to:
1. Ability to perform direct expansion of variables in a string
2. Ability to specify default values via a prioritized mapping function but without support for
defaults as a syntax-level feature
1. Ability to perform direct expansion of variables in a string.
2. Ability to specify default values via a prioritized mapping function but
without support for defaults as a syntax-level feature.
### What should the syntax be?
The exact syntax for variable expansion has a large impact on how users perceive and relate to the
feature. We considered implementing a very restrictive subset of the shell `${var}` syntax. This
syntax is an attractive option on some level, because many people are familiar with it. However,
this syntax also has a large number of lesser known features such as the ability to provide
default values for unset variables, perform inline substitution, etc.
The exact syntax for variable expansion has a large impact on how users perceive
and relate to the feature. We considered implementing a very restrictive subset
of the shell `${var}` syntax. This syntax is an attractive option on some level,
because many people are familiar with it. However, this syntax also has a large
number of lesser known features such as the ability to provide default values
for unset variables, perform inline substitution, etc.
In the interest of preventing conflation of the expansion feature in Kubernetes with the shell
feature, we chose a different syntax similar to the one in Makefiles, `$(var)`. We also chose not
to support the bar `$var` format, since it is not required to implement the required use-cases.
In the interest of preventing conflation of the expansion feature in Kubernetes
with the shell feature, we chose a different syntax similar to the one in
Makefiles, `$(var)`. We also chose not to support the bar `$var` format, since
it is not required to implement the required use-cases.
Nested references, ie, variable expansion within variable names, are not supported.
Nested references, ie, variable expansion within variable names, are not
supported.
#### How should unmatched references be treated?
Ideally, it should be extremely clear when a variable reference couldn't be expanded. We decided
the best experience for unmatched variable references would be to have the entire reference, syntax
included, show up in the output. As an example, if the reference `$(VARIABLE_NAME)` cannot be
expanded, then `$(VARIABLE_NAME)` should be present in the output.
Ideally, it should be extremely clear when a variable reference couldn't be
expanded. We decided the best experience for unmatched variable references would
be to have the entire reference, syntax included, show up in the output. As an
example, if the reference `$(VARIABLE_NAME)` cannot be expanded, then
`$(VARIABLE_NAME)` should be present in the output.
#### Escaping the operator
Although the `$(var)` syntax does overlap with the `$(command)` form of command substitution
supported by many shells, because unexpanded variables are present verbatim in the output, we
expect this will not present a problem to many users. If there is a collision between a variable
name and command substitution syntax, the syntax can be escaped with the form `$$(VARIABLE_NAME)`,
which will evaluate to `$(VARIABLE_NAME)` whether `VARIABLE_NAME` can be expanded or not.
Although the `$(var)` syntax does overlap with the `$(command)` form of command
substitution supported by many shells, because unexpanded variables are present
verbatim in the output, we expect this will not present a problem to many users.
If there is a collision between a variable name and command substitution syntax,
the syntax can be escaped with the form `$$(VARIABLE_NAME)`, which will evaluate
to `$(VARIABLE_NAME)` whether `VARIABLE_NAME` can be expanded or not.
## Design
This design encompasses the variable expansion syntax and specification and the changes needed to
incorporate the expansion feature into the container's environment and command.
This design encompasses the variable expansion syntax and specification and the
changes needed to incorporate the expansion feature into the container's
environment and command.
### Syntax and expansion mechanics
This section describes the expansion syntax, evaluation of variable values, and how unexpected or
malformed inputs are handled.
This section describes the expansion syntax, evaluation of variable values, and
how unexpected or malformed inputs are handled.
#### Syntax
The inputs to the expansion feature are:
1. A utf-8 string (the input string) which may contain variable references
2. A function (the mapping function) that maps the name of a variable to the variable's value, of
type `func(string) string`
1. A utf-8 string (the input string) which may contain variable references.
2. A function (the mapping function) that maps the name of a variable to the
variable's value, of type `func(string) string`.
Variable references in the input string are indicated exclusively with the syntax
`$(<variable-name>)`. The syntax tokens are:
`$(<variable-name>)`. The syntax tokens are:
- `$`: the operator
- `(`: the reference opener
- `)`: the reference closer
- `$`: the operator,
- `(`: the reference opener, and
- `)`: the reference closer.
The operator has no meaning unless accompanied by the reference opener and closer tokens. The
operator can be escaped using `$$`. One literal `$` will be emitted for each `$$` in the input.
The operator has no meaning unless accompanied by the reference opener and
closer tokens. The operator can be escaped using `$$`. One literal `$` will be
emitted for each `$$` in the input.
The reference opener and closer characters have no meaning when not part of a variable reference.
If a variable reference is malformed, viz: `$(VARIABLE_NAME` without a closing expression, the
operator and expression opening characters are treated as ordinary characters without special
meanings.
The reference opener and closer characters have no meaning when not part of a
variable reference. If a variable reference is malformed, viz: `$(VARIABLE_NAME`
without a closing expression, the operator and expression opening characters are
treated as ordinary characters without special meanings.
#### Scope and ordering of substitutions
The scope in which variable references are expanded is defined by the mapping function. Within the
mapping function, any arbitrary strategy may be used to determine the value of a variable name.
The most basic implementation of a mapping function is to use a `map[string]string` to lookup the
value of a variable.
The scope in which variable references are expanded is defined by the mapping
function. Within the mapping function, any arbitrary strategy may be used to
determine the value of a variable name. The most basic implementation of a
mapping function is to use a `map[string]string` to lookup the value of a
variable.
In order to support default values for variables like service variables presented by the kubelet,
which may not be bound because the service that provides them does not yet exist, there should be a
mapping function that uses a list of `map[string]string` like:
In order to support default values for variables like service variables
presented by the kubelet, which may not be bound because the service that
provides them does not yet exist, there should be a mapping function that uses a
list of `map[string]string` like:
```go
func MakeMappingFunc(maps ...map[string]string) func(string) string {
@ -235,38 +257,41 @@ mappingWithDefaults := MakeMappingFunc(serviceEnv, containerEnv)
The necessary changes to implement this functionality are:
1. Add a new interface, `ObjectEventRecorder`, which is like the `EventRecorder` interface, but
scoped to a single object, and a function that returns an `ObjectEventRecorder` given an
`ObjectReference` and an `EventRecorder`
1. Add a new interface, `ObjectEventRecorder`, which is like the
`EventRecorder` interface, but scoped to a single object, and a function that
returns an `ObjectEventRecorder` given an `ObjectReference` and an
`EventRecorder`.
2. Introduce `third_party/golang/expansion` package that provides:
1. An `Expand(string, func(string) string) string` function
2. A `MappingFuncFor(ObjectEventRecorder, ...map[string]string) string` function
3. Make the kubelet expand environment correctly
4. Make the kubelet expand command correctly
1. An `Expand(string, func(string) string) string` function.
2. A `MappingFuncFor(ObjectEventRecorder, ...map[string]string) string`
function.
3. Make the kubelet expand environment correctly.
4. Make the kubelet expand command correctly.
#### Event Recording
In order to provide an event when an expansion references undefined variables, the mapping function
must be able to create an event. In order to facilitate this, we should create a new interface in
the `api/client/record` package which is similar to `EventRecorder`, but scoped to a single object:
In order to provide an event when an expansion references undefined variables,
the mapping function must be able to create an event. In order to facilitate
this, we should create a new interface in the `api/client/record` package which
is similar to `EventRecorder`, but scoped to a single object:
```go
// ObjectEventRecorder knows how to record events about a single object.
type ObjectEventRecorder interface {
// Event constructs an event from the given information and puts it in the queue for sending.
// 'reason' is the reason this event is generated. 'reason' should be short and unique; it will
// be used to automate handling of events, so imagine people writing switch statements to
// handle them. You want to make that easy.
// 'message' is intended to be human readable.
//
// The resulting event will be created in the same namespace as the reference object.
Event(reason, message string)
// Event constructs an event from the given information and puts it in the queue for sending.
// 'reason' is the reason this event is generated. 'reason' should be short and unique; it will
// be used to automate handling of events, so imagine people writing switch statements to
// handle them. You want to make that easy.
// 'message' is intended to be human readable.
//
// The resulting event will be created in the same namespace as the reference object.
Event(reason, message string)
// Eventf is just like Event, but with Sprintf for the message field.
Eventf(reason, messageFmt string, args ...interface{})
// Eventf is just like Event, but with Sprintf for the message field.
Eventf(reason, messageFmt string, args ...interface{})
// PastEventf is just like Eventf, but with an option to specify the event's 'timestamp' field.
PastEventf(timestamp unversioned.Time, reason, messageFmt string, args ...interface{})
// PastEventf is just like Eventf, but with an option to specify the event's 'timestamp' field.
PastEventf(timestamp unversioned.Time, reason, messageFmt string, args ...interface{})
}
```
@ -275,16 +300,16 @@ and an `EventRecorder`:
```go
type objectRecorderImpl struct {
object runtime.Object
recorder EventRecorder
object runtime.Object
recorder EventRecorder
}
func (r *objectRecorderImpl) Event(reason, message string) {
r.recorder.Event(r.object, reason, message)
r.recorder.Event(r.object, reason, message)
}
func ObjectEventRecorderFor(object runtime.Object, recorder EventRecorder) ObjectEventRecorder {
return &objectRecorderImpl{object, recorder}
return &objectRecorderImpl{object, recorder}
}
```
@ -299,28 +324,29 @@ The expansion package should provide two methods:
// for the input is found. If no expansion is found for a key, an event
// is raised on the given recorder.
func MappingFuncFor(recorder record.ObjectEventRecorder, context ...map[string]string) func(string) string {
// ...
// ...
}
// Expand replaces variable references in the input string according to
// the expansion spec using the given mapping function to resolve the
// values of variables.
func Expand(input string, mapping func(string) string) string {
// ...
// ...
}
```
#### Kubelet changes
The Kubelet should be made to correctly expand variables references in a container's environment,
command, and args. Changes will need to be made to:
The Kubelet should be made to correctly expand variables references in a
container's environment, command, and args. Changes will need to be made to:
1. The `makeEnvironmentVariables` function in the kubelet; this is used by
`GenerateRunContainerOptions`, which is used by both the docker and rkt container runtimes
2. The docker manager `setEntrypointAndCommand` func has to be changed to perform variable
expansion
3. The rkt runtime should be made to support expansion in command and args when support for it is
implemented
`GenerateRunContainerOptions`, which is used by both the docker and rkt
container runtimes.
2. The docker manager `setEntrypointAndCommand` func has to be changed to
perform variable expansion.
3. The rkt runtime should be made to support expansion in command and args
when support for it is implemented.
### Examples

109
docs/design/extending-api.md
View File

@ -34,59 +34,62 @@ Documentation for other releases can be found at
# Adding custom resources to the Kubernetes API server
This document describes the design for implementing the storage of custom API types in the Kubernetes API Server.
This document describes the design for implementing the storage of custom API
types in the Kubernetes API Server.
## Resource Model
### The ThirdPartyResource
The `ThirdPartyResource` resource describes the multiple versions of a custom resource that the user wants to add
to the Kubernetes API. `ThirdPartyResource` is a non-namespaced resource; attempting to place it in a namespace
will return an error.
The `ThirdPartyResource` resource describes the multiple versions of a custom
resource that the user wants to add to the Kubernetes API. `ThirdPartyResource`
is a non-namespaced resource; attempting to place it in a namespace will return
an error.
Each `ThirdPartyResource` resource has the following:
* Standard Kubernetes object metadata.
* ResourceKind - The kind of the resources described by this third party resource.
* ResourceKind - The kind of the resources described by this third party
resource.
* Description - A free text description of the resource.
* APIGroup - An API group that this resource should be placed into.
* Versions - One or more `Version` objects.
### The `Version` Object
The `Version` object describes a single concrete version of a custom resource. The `Version` object currently
only specifies:
The `Version` object describes a single concrete version of a custom resource.
The `Version` object currently only specifies:
* The `Name` of the version.
* The `APIGroup` this version should belong to.
## Expectations about third party objects
Every object that is added to a third-party Kubernetes object store is expected to contain Kubernetes
compatible [object metadata](../devel/api-conventions.md#metadata). This requirement enables the
Kubernetes API server to provide the following features:
* Filtering lists of objects via label queries
* `resourceVersion`-based optimistic concurrency via compare-and-swap
* Versioned storage
* Event recording
* Integration with basic `kubectl` command line tooling
* Watch for resource changes
Every object that is added to a third-party Kubernetes object store is expected
to contain Kubernetes compatible [object metadata](../devel/api-conventions.md#metadata).
This requirement enables the Kubernetes API server to provide the following
features:
* Filtering lists of objects via label queries.
* `resourceVersion`-based optimistic concurrency via compare-and-swap.
* Versioned storage.
* Event recording.
* Integration with basic `kubectl` command line tooling.
* Watch for resource changes.
The `Kind` for an instance of a third-party object (e.g. CronTab) below is expected to be
programmatically convertible to the name of the resource using
the following conversion. Kinds are expected to be of the form `<CamelCaseKind>`, and the
`APIVersion` for the object is expected to be `<api-group>/<api-version>`. To
prevent collisions, it's expected that you'll use a fully qualified domain
name for the API group, e.g. `example.com`.
The `Kind` for an instance of a third-party object (e.g. CronTab) below is
expected to be programmatically convertible to the name of the resource using
the following conversion. Kinds are expected to be of the form
`<CamelCaseKind>`, and the `APIVersion` for the object is expected to be
`<api-group>/<api-version>`. To prevent collisions, it's expected that you'll
use a fully qualified domain name for the API group, e.g. `example.com`.
For example `stable.example.com/v1`
'CamelCaseKind' is the specific type name.
To convert this into the `metadata.name` for the `ThirdPartyResource` resource instance,
the `<domain-name>` is copied verbatim, the `CamelCaseKind` is
then converted
using '-' instead of capitalization ('camel-case'), with the first character being assumed to be
capitalized. In pseudo code:
To convert this into the `metadata.name` for the `ThirdPartyResource` resource
instance, the `<domain-name>` is copied verbatim, the `CamelCaseKind` is then
converted using '-' instead of capitalization ('camel-case'), with the first
character being assumed to be capitalized. In pseudo code:
```go
var result string
@ -98,17 +101,20 @@ for ix := range kindName {
}
```
As a concrete example, the resource named `camel-case-kind.example.com` defines resources of Kind `CamelCaseKind`, in
the APIGroup with the prefix `example.com/...`.
As a concrete example, the resource named `camel-case-kind.example.com` defines
resources of Kind `CamelCaseKind`, in the APIGroup with the prefix
`example.com/...`.
The reason for this is to enable rapid lookup of a `ThirdPartyResource` object given the kind information.
This is also the reason why `ThirdPartyResource` is not namespaced.
The reason for this is to enable rapid lookup of a `ThirdPartyResource` object
given the kind information. This is also the reason why `ThirdPartyResource` is
not namespaced.
## Usage
When a user creates a new `ThirdPartyResource`, the Kubernetes API Server reacts by creating a new, namespaced
RESTful resource path. For now, non-namespaced objects are not supported. As with existing built-in objects,
deleting a namespace deletes all third party resources in that namespace.
When a user creates a new `ThirdPartyResource`, the Kubernetes API Server reacts
by creating a new, namespaced RESTful resource path. For now, non-namespaced
objects are not supported. As with existing built-in objects, deleting a
namespace deletes all third party resources in that namespace.
For example, if a user creates:
@ -143,14 +149,15 @@ Now that this schema has been created, a user can `POST`:
to: `/apis/stable.example.com/v1/namespaces/default/crontabs`
and the corresponding data will be stored into etcd by the APIServer, so that when the user issues:
and the corresponding data will be stored into etcd by the APIServer, so that
when the user issues:
```
GET /apis/stable.example.com/v1/namespaces/default/crontabs/my-new-cron-object`
```
And when they do that, they will get back the same data, but with additional Kubernetes metadata
(e.g. `resourceVersion`, `createdTimestamp`) filled in.
And when they do that, they will get back the same data, but with additional
Kubernetes metadata (e.g. `resourceVersion`, `createdTimestamp`) filled in.
Likewise, to list all resources, a user can issue:
@ -178,29 +185,35 @@ and get back:
}
```
Because all objects are expected to contain standard Kubernetes metadata fields, these
list operations can also use label queries to filter requests down to specific subsets.
Likewise, clients can use watch endpoints to watch for changes to stored objects.
Because all objects are expected to contain standard Kubernetes metadata fields,
these list operations can also use label queries to filter requests down to
specific subsets.
Likewise, clients can use watch endpoints to watch for changes to stored
objects.
## Storage
In order to store custom user data in a versioned fashion inside of etcd, we need to also introduce a
`Codec`-compatible object for persistent storage in etcd. This object is `ThirdPartyResourceData` and it contains:
* Standard API Metadata
In order to store custom user data in a versioned fashion inside of etcd, we
need to also introduce a `Codec`-compatible object for persistent storage in
etcd. This object is `ThirdPartyResourceData` and it contains:
* Standard API Metadata.
* `Data`: The raw JSON data for this custom object.
### Storage key specification
Each custom object stored by the API server needs a custom key in storage, this is described below:
Each custom object stored by the API server needs a custom key in storage, this
is described below:
#### Definitions
* `resource-namespace`: the namespace of the particular resource that is being stored
* `resource-namespace`: the namespace of the particular resource that is
being stored
* `resource-name`: the name of the particular resource being stored
* `third-party-resource-namespace`: the namespace of the `ThirdPartyResource` resource that represents the type for the specific instance being stored
* `third-party-resource-name`: the name of the `ThirdPartyResource` resource that represents the type for the specific instance being stored
* `third-party-resource-namespace`: the namespace of the `ThirdPartyResource`
resource that represents the type for the specific instance being stored
* `third-party-resource-name`: the name of the `ThirdPartyResource` resource
that represents the type for the specific instance being stored
#### Key

173
docs/design/federated-services.md
View File

@ -76,7 +76,7 @@ Documentation for other releases can be found at
load balancers between the client and the serving Pod, failover
might be completely automatic (i.e. the client's end of the
connection remains intact, and the client is completely
oblivious of the fail-over). This approach incurs network speed
oblivious of the fail-over). This approach incurs network speed
and cost penalties (by traversing possibly multiple load
balancers), but requires zero smarts in clients, DNS libraries,
recursing DNS servers etc, as the IP address of the endpoint
@ -102,17 +102,17 @@ Documentation for other releases can be found at
A Kubernetes application configuration (e.g. for a Pod, Replication
Controller, Service etc) should be able to be successfully deployed
into any Kubernetes Cluster or Ubernetes Federation of Clusters,
without modification. More specifically, a typical configuration
without modification. More specifically, a typical configuration
should work correctly (although possibly not optimally) across any of
the following environments:
1. A single Kubernetes Cluster on one cloud provider (e.g. Google
Compute Engine, GCE)
Compute Engine, GCE).
1. A single Kubernetes Cluster on a different cloud provider
(e.g. Amazon Web Services, AWS)
(e.g. Amazon Web Services, AWS).
1. A single Kubernetes Cluster on a non-cloud, on-premise data center
1. A Federation of Kubernetes Clusters all on the same cloud provider
(e.g. GCE)
(e.g. GCE).
1. A Federation of Kubernetes Clusters across multiple different cloud
providers and/or on-premise data centers (e.g. one cluster on
GCE/GKE, one on AWS, and one on-premise).
@ -122,18 +122,18 @@ the following environments:
It should be possible to explicitly opt out of portability across some
subset of the above environments in order to take advantage of
non-portable load balancing and DNS features of one or more
environments. More specifically, for example:
environments. More specifically, for example:
1. For HTTP(S) applications running on GCE-only Federations,
[GCE Global L7 Load Balancers](https://cloud.google.com/compute/docs/load-balancing/http/global-forwarding-rules)
should be usable. These provide single, static global IP addresses
should be usable. These provide single, static global IP addresses
which load balance and fail over globally (i.e. across both regions
and zones). These allow for really dumb clients, but they only
and zones). These allow for really dumb clients, but they only
work on GCE, and only for HTTP(S) traffic.
1. For non-HTTP(S) applications running on GCE-only Federations within
a single region,
[GCE L4 Network Load Balancers](https://cloud.google.com/compute/docs/load-balancing/network/)
should be usable. These provide TCP (i.e. both HTTP/S and
should be usable. These provide TCP (i.e. both HTTP/S and
non-HTTP/S) load balancing and failover, but only on GCE, and only
within a single region.
[Google Cloud DNS](https://cloud.google.com/dns) can be used to
@ -141,7 +141,7 @@ environments. More specifically, for example:
providers and on-premise clusters, as it's plain DNS, IP only).
1. For applications running on AWS-only Federations,
[AWS Elastic Load Balancers (ELB's)](https://aws.amazon.com/elasticloadbalancing/details/)
should be usable. These provide both L7 (HTTP(S)) and L4 load
should be usable. These provide both L7 (HTTP(S)) and L4 load
balancing, but only within a single region, and only on AWS
([AWS Route 53 DNS service](https://aws.amazon.com/route53/) can be
used to load balance and fail over across multiple regions, and is
@ -153,7 +153,7 @@ Ubernetes cross-cluster load balancing is built on top of the following:
1. [GCE Global L7 Load Balancers](https://cloud.google.com/compute/docs/load-balancing/http/global-forwarding-rules)
provide single, static global IP addresses which load balance and
fail over globally (i.e. across both regions and zones). These
fail over globally (i.e. across both regions and zones). These
allow for really dumb clients, but they only work on GCE, and only
for HTTP(S) traffic.
1. [GCE L4 Network Load Balancers](https://cloud.google.com/compute/docs/load-balancing/network/)
@ -170,7 +170,7 @@ Ubernetes cross-cluster load balancing is built on top of the following:
doesn't provide any built-in geo-DNS, latency-based routing, health
checking, weighted round robin or other advanced capabilities.
It's plain old DNS. We would need to build all the aforementioned
on top of it. It can provide internal DNS services (i.e. serve RFC
on top of it. It can provide internal DNS services (i.e. serve RFC
1918 addresses).
1. [AWS Route 53 DNS service](https://aws.amazon.com/route53/) can
be used to load balance and fail over across regions, and is also
@ -185,23 +185,24 @@ Ubernetes cross-cluster load balancing is built on top of the following:
service IP which is load-balanced (currently simple round-robin)
across the healthy pods comprising a service within a single
Kubernetes cluster.
1. [Kubernetes Ingress](http://kubernetes.io/v1.1/docs/user-guide/ingress.html): A generic wrapper around cloud-provided L4 and L7 load balancing services, and roll-your-own load balancers run in pods, e.g. HA Proxy.
1. [Kubernetes Ingress](http://kubernetes.io/v1.1/docs/user-guide/ingress.html):
A generic wrapper around cloud-provided L4 and L7 load balancing services, and
roll-your-own load balancers run in pods, e.g. HA Proxy.
## Ubernetes API
The Ubernetes API for load balancing should be compatible with the
equivalent Kubernetes API, to ease porting of clients between
Ubernetes and Kubernetes. Further details below.
The Ubernetes API for load balancing should be compatible with the equivalent
Kubernetes API, to ease porting of clients between Ubernetes and Kubernetes.
Further details below.
## Common Client Behavior
To be useful, our load balancing solution needs to work properly with
real client applications. There are a few different classes of
those...
To be useful, our load balancing solution needs to work properly with real
client applications. There are a few different classes of those...
### Browsers
These are the most common external clients. These are all well-written. See below.
These are the most common external clients. These are all well-written. See below.
### Well-written clients
@ -218,8 +219,8 @@ Examples:
### Dumb clients
1. Don't do a DNS resolution every time they connect (or do cache
beyond the TTL).
1. Don't do a DNS resolution every time they connect (or do cache beyond the
TTL).
1. Do try multiple A records
Examples:
@ -237,34 +238,34 @@ Examples:
### Dumbest clients
1. Never do a DNS lookup - are pre-configured with a single (or
possibly multiple) fixed server IP(s). Nothing else matters.
1. Never do a DNS lookup - are pre-configured with a single (or possibly
multiple) fixed server IP(s). Nothing else matters.
## Architecture and Implementation
### General control plane architecture
### General Control Plane Architecture
Each cluster hosts one or more Ubernetes master components (Ubernetes API servers, controller managers with leader election, and
etcd quorum members. This is documented in more detail in a
[separate design doc: Kubernetes/Ubernetes Control Plane Resilience](https://docs.google.com/document/d/1jGcUVg9HDqQZdcgcFYlWMXXdZsplDdY6w3ZGJbU7lAw/edit#).
Each cluster hosts one or more Ubernetes master components (Ubernetes API
servers, controller managers with leader election, and etcd quorum members. This
is documented in more detail in a separate design doc:
[Kubernetes/Ubernetes Control Plane Resilience](https://docs.google.com/document/d/1jGcUVg9HDqQZdcgcFYlWMXXdZsplDdY6w3ZGJbU7lAw/edit#).
In the description below, assume that 'n' clusters, named
'cluster-1'... 'cluster-n' have been registered against an Ubernetes
Federation "federation-1", each with their own set of Kubernetes API
endpoints,so,
In the description below, assume that 'n' clusters, named 'cluster-1'...
'cluster-n' have been registered against an Ubernetes Federation "federation-1",
each with their own set of Kubernetes API endpoints,so,
"[http://endpoint-1.cluster-1](http://endpoint-1.cluster-1),
[http://endpoint-2.cluster-1](http://endpoint-2.cluster-1)
... [http://endpoint-m.cluster-n](http://endpoint-m.cluster-n) .
### Federated Services
Ubernetes Services are pretty straight-forward. They're comprised of
multiple equivalent underlying Kubernetes Services, each with their
own external endpoint, and a load balancing mechanism across them.
Let's work through how exactly that works in practice.
Ubernetes Services are pretty straight-forward. They're comprised of multiple
equivalent underlying Kubernetes Services, each with their own external
endpoint, and a load balancing mechanism across them. Let's work through how
exactly that works in practice.
Our user creates the following Ubernetes Service (against an Ubernetes
API endpoint):
Our user creates the following Ubernetes Service (against an Ubernetes API
endpoint):
$ kubectl create -f my-service.yaml --context="federation-1"
@ -290,9 +291,9 @@ where service.yaml contains the following:
run: my-service
type: LoadBalancer
Ubernetes in turn creates one equivalent service (identical config to
the above) in each of the underlying Kubernetes clusters, each of
which results in something like this:
Ubernetes in turn creates one equivalent service (identical config to the above)
in each of the underlying Kubernetes clusters, each of which results in
something like this:
$ kubectl get -o yaml --context="cluster-1" service my-service
@ -329,9 +330,8 @@ which results in something like this:
ingress:
- ip: 104.197.117.10
Similar services are created in `cluster-2` and `cluster-3`, each of
which are allocated their own `spec.clusterIP`, and
`status.loadBalancer.ingress.ip`.
Similar services are created in `cluster-2` and `cluster-3`, each of which are
allocated their own `spec.clusterIP`, and `status.loadBalancer.ingress.ip`.
In Ubernetes `federation-1`, the resulting federated service looks as follows:
@ -376,21 +376,21 @@ Note that the federated service:
1. has no clusterIP (as it is cluster-independent)
1. has a federation-wide load balancer hostname
In addition to the set of underlying Kubernetes services (one per
cluster) described above, Ubernetes has also created a DNS name
(e.g. on [Google Cloud DNS](https://cloud.google.com/dns) or
[AWS Route 53](https://aws.amazon.com/route53/), depending on
configuration) which provides load balancing across all of those
services. For example, in a very basic configuration:
In addition to the set of underlying Kubernetes services (one per cluster)
described above, Ubernetes has also created a DNS name (e.g. on
[Google Cloud DNS](https://cloud.google.com/dns) or
[AWS Route 53](https://aws.amazon.com/route53/), depending on configuration)
which provides load balancing across all of those services. For example, in a
very basic configuration:
$ dig +noall +answer my-service.my-namespace.my-federation.my-domain.com
my-service.my-namespace.my-federation.my-domain.com 180 IN A 104.197.117.10
my-service.my-namespace.my-federation.my-domain.com 180 IN A 104.197.74.77
my-service.my-namespace.my-federation.my-domain.com 180 IN A 104.197.38.157
Each of the above IP addresses (which are just the external load
balancer ingress IP's of each cluster service) is of course load
balanced across the pods comprising the service in each cluster.
Each of the above IP addresses (which are just the external load balancer
ingress IP's of each cluster service) is of course load balanced across the pods
comprising the service in each cluster.
In a more sophisticated configuration (e.g. on GCE or GKE), Ubernetes
automatically creates a
@ -411,23 +411,21 @@ for failover purposes:
my-service.my-namespace.my-federation.my-domain.com 180 IN A 104.197.74.77
my-service.my-namespace.my-federation.my-domain.com 180 IN A 104.197.38.157
If Ubernetes Global Service Health Checking is enabled, multiple
service health checkers running across the federated clusters
collaborate to monitor the health of the service endpoints, and
automatically remove unhealthy endpoints from the DNS record (e.g. a
majority quorum is required to vote a service endpoint unhealthy, to
avoid false positives due to individual health checker network
If Ubernetes Global Service Health Checking is enabled, multiple service health
checkers running across the federated clusters collaborate to monitor the health
of the service endpoints, and automatically remove unhealthy endpoints from the
DNS record (e.g. a majority quorum is required to vote a service endpoint
unhealthy, to avoid false positives due to individual health checker network
isolation).
### Federated Replication Controllers
So far we have a federated service defined, with a resolvable load
balancer hostname by which clients can reach it, but no pods serving
traffic directed there. So now we need a Federated Replication
Controller. These are also fairly straight-forward, being comprised
of multiple underlying Kubernetes Replication Controllers which do the
hard work of keeping the desired number of Pod replicas alive in each
Kubernetes cluster.
So far we have a federated service defined, with a resolvable load balancer
hostname by which clients can reach it, but no pods serving traffic directed
there. So now we need a Federated Replication Controller. These are also fairly
straight-forward, being comprised of multiple underlying Kubernetes Replication
Controllers which do the hard work of keeping the desired number of Pod replicas
alive in each Kubernetes cluster.
$ kubectl create -f my-service-rc.yaml --context="federation-1"
@ -495,54 +493,49 @@ something like this:
status:
replicas: 2
The exact number of replicas created in each underlying cluster will
of course depend on what scheduling policy is in force. In the above
example, the scheduler created an equal number of replicas (2) in each
of the three underlying clusters, to make up the total of 6 replicas
required. To handle entire cluster failures, various approaches are possible,
including:
The exact number of replicas created in each underlying cluster will of course
depend on what scheduling policy is in force. In the above example, the
scheduler created an equal number of replicas (2) in each of the three
underlying clusters, to make up the total of 6 replicas required. To handle
entire cluster failures, various approaches are possible, including:
1. **simple overprovisioing**, such that sufficient replicas remain even if a
cluster fails. This wastes some resources, but is simple and
reliable.
cluster fails. This wastes some resources, but is simple and reliable.
2. **pod autoscaling**, where the replication controller in each
cluster automatically and autonomously increases the number of
replicas in its cluster in response to the additional traffic
diverted from the
failed cluster. This saves resources and is reatively simple,
but there is some delay in the autoscaling.
diverted from the failed cluster. This saves resources and is relatively
simple, but there is some delay in the autoscaling.
3. **federated replica migration**, where the Ubernetes Federation
Control Plane detects the cluster failure and automatically
increases the replica count in the remainaing clusters to make up
for the lost replicas in the failed cluster. This does not seem to
for the lost replicas in the failed cluster. This does not seem to
offer any benefits relative to pod autoscaling above, and is
arguably more complex to implement, but we note it here as a
possibility.
### Implementation Details
The implementation approach and architecture is very similar to
Kubernetes, so if you're familiar with how Kubernetes works, none of
what follows will be surprising. One additional design driver not
present in Kubernetes is that Ubernetes aims to be resilient to
individual cluster and availability zone failures. So the control
plane spans multiple clusters. More specifically:
The implementation approach and architecture is very similar to Kubernetes, so
if you're familiar with how Kubernetes works, none of what follows will be
surprising. One additional design driver not present in Kubernetes is that
Ubernetes aims to be resilient to individual cluster and availability zone
failures. So the control plane spans multiple clusters. More specifically:
+ Ubernetes runs it's own distinct set of API servers (typically one
or more per underlying Kubernetes cluster). These are completely
distinct from the Kubernetes API servers for each of the underlying
clusters.
+ Ubernetes runs it's own distinct quorum-based metadata store (etcd,
by default). Approximately 1 quorum member runs in each underlying
by default). Approximately 1 quorum member runs in each underlying
cluster ("approximately" because we aim for an odd number of quorum
members, and typically don't want more than 5 quorum members, even
if we have a larger number of federated clusters, so 2 clusters->3
quorum members, 3->3, 4->3, 5->5, 6->5, 7->5 etc).
Cluster Controllers in Ubernetes watch against the Ubernetes API
server/etcd state, and apply changes to the underlying kubernetes
clusters accordingly. They also have the anti-entropy mechanism for
reconciling ubernetes "desired desired" state against kubernetes
"actual desired" state.
Cluster Controllers in Ubernetes watch against the Ubernetes API server/etcd
state, and apply changes to the underlying kubernetes clusters accordingly. They
also have the anti-entropy mechanism for reconciling ubernetes "desired desired"
state against kubernetes "actual desired" state.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

38
docs/design/federation-phase-1.md
View File

@ -71,7 +71,8 @@ unified view.
Here are the functionality requirements derived from above use cases:
+ Clients of the federation control plane API server can register and deregister clusters.
+ Clients of the federation control plane API server can register and deregister
clusters.
+ Workloads should be spread to different clusters according to the
workload distribution policy.
+ Pods are able to discover and connect to services hosted in other
@ -90,7 +91,7 @@ Here are the functionality requirements derived from above use cases:
It’s difficult to have a perfect design with one click that implements
all the above requirements. Therefore we will go with an iterative
approach to design and build the system. This document describes the
phase one of the whole work. In phase one we will cover only the
phase one of the whole work. In phase one we will cover only the
following objectives:
+ Define the basic building blocks and API objects of control plane
@ -130,9 +131,9 @@ description of each module contained in above diagram.
The API Server in the Ubernetes control plane works just like the API
Server in K8S. It talks to a distributed key-value store to persist,
retrieve and watch API objects. This store is completely distinct
retrieve and watch API objects. This store is completely distinct
from the kubernetes key-value stores (etcd) in the underlying
kubernetes clusters. We still use `etcd` as the distributed
kubernetes clusters. We still use `etcd` as the distributed
storage so customers don’t need to learn and manage a different
storage system, although it is envisaged that other storage systems
(consol, zookeeper) will probably be developedand supported over
@ -141,16 +142,16 @@ time.
## Ubernetes Scheduler
The Ubernetes Scheduler schedules resources onto the underlying
Kubernetes clusters. For example it watches for unscheduled Ubernetes
Kubernetes clusters. For example it watches for unscheduled Ubernetes
replication controllers (those that have not yet been scheduled onto
underlying Kubernetes clusters) and performs the global scheduling
work. For each unscheduled replication controller, it calls policy
work. For each unscheduled replication controller, it calls policy
engine to decide how to spit workloads among clusters. It creates a
Kubernetes Replication Controller on one ore more underlying cluster,
and post them back to `etcd` storage.
One sublety worth noting here is that the scheduling decision is
arrived at by combining the application-specific request from the user (which might
One sublety worth noting here is that the scheduling decision is arrived at by
combining the application-specific request from the user (which might
include, for example, placement constraints), and the global policy specified
by the federation administrator (for example, "prefer on-premise
clusters over AWS clusters" or "spread load equally across clusters").
@ -165,9 +166,9 @@ performs the following two kinds of work:
corresponding API objects on the underlying K8S clusters.
1. It periodically retrieves the available resources metrics from the
underlying K8S cluster, and updates them as object status of the
`cluster` API object. An alternative design might be to run a pod
`cluster` API object. An alternative design might be to run a pod
in each underlying cluster that reports metrics for that cluster to
the Ubernetes control plane. Which approach is better remains an
the Ubernetes control plane. Which approach is better remains an
open topic of discussion.
## Ubernetes Service Controller
@ -187,7 +188,7 @@ Cluster is a new first-class API object introduced in this design. For
each registered K8S cluster there will be such an API resource in
control plane. The way clients register or deregister a cluster is to
send corresponding REST requests to following URL:
`/api/{$version}/clusters`. Because control plane is behaving like a
`/api/{$version}/clusters`. Because control plane is behaving like a
regular K8S client to the underlying clusters, the spec of a cluster
object contains necessary properties like K8S cluster address and
credentials. The status of a cluster API object will contain
@ -294,7 +295,7 @@ $version.clusterStatus
**For simplicity we didn’t introduce a separate “cluster metrics” API
object here**. The cluster resource metrics are stored in cluster
status section, just like what we did to nodes in K8S. In phase one it
only contains available CPU resources and memory resources. The
only contains available CPU resources and memory resources. The
cluster controller will periodically poll the underlying cluster API
Server to get cluster capability. In phase one it gets the metrics by
simply aggregating metrics from all nodes. In future we will improve
@ -315,7 +316,7 @@ Below is the state transition diagram.
## Replication Controller
A global workload submitted to control plane is represented as an
Ubernetes replication controller. When a replication controller
Ubernetes replication controller. When a replication controller
is submitted to control plane, clients need a way to express its
requirements or preferences on clusters. Depending on different use
cases it may be complex. For example:
@ -327,7 +328,7 @@ cases it may be complex. For example:
(use case: workload )
+ Seventy percent of this workload should be scheduled to cluster Foo,
and thirty percent should be scheduled to cluster Bar (use case:
vendor lock-in avoidance). In phase one, we only introduce a
vendor lock-in avoidance). In phase one, we only introduce a
_clusterSelector_ field to filter acceptable clusters. In default
case there is no such selector and it means any cluster is
acceptable.
@ -376,7 +377,7 @@ clusters. How to handle this will be addressed after phase one.
The Service API object exposed by Ubernetes is similar to service
objects on Kubernetes. It defines the access to a group of pods. The
Ubernetes service controller will create corresponding Kubernetes
service objects on underlying clusters. These are detailed in a
service objects on underlying clusters. These are detailed in a
separate design document: [Federated Services](federated-services.md).
## Pod
@ -389,7 +390,8 @@ order to keep the Ubernetes API compatible with the Kubernetes API.
## Scheduling
The below diagram shows how workloads are scheduled on the Ubernetes control plane:
The below diagram shows how workloads are scheduled on the Ubernetes control\
plane:
1. A replication controller is created by the client.
1. APIServer persists it into the storage.
@ -425,8 +427,8 @@ proposed solutions like resource reservation mechanisms.
This part has been included in the section “Federated Service” of
document
“[Ubernetes Cross-cluster Load Balancing and Service Discovery Requirements and System Design](federated-services.md))”. Please
refer to that document for details.
“[Ubernetes Cross-cluster Load Balancing and Service Discovery Requirements and System Design](federated-services.md))”.
Please refer to that document for details.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

182
docs/design/horizontal-pod-autoscaler.md
View File

@ -36,33 +36,40 @@ Documentation for other releases can be found at
## Preface
This document briefly describes the design of the horizontal autoscaler for pods.
The autoscaler (implemented as a Kubernetes API resource and controller) is responsible for dynamically controlling
the number of replicas of some collection (e.g. the pods of a ReplicationController) to meet some objective(s),
This document briefly describes the design of the horizontal autoscaler for
pods. The autoscaler (implemented as a Kubernetes API resource and controller)
is responsible for dynamically controlling the number of replicas of some
collection (e.g. the pods of a ReplicationController) to meet some objective(s),
for example a target per-pod CPU utilization.
This design supersedes [autoscaling.md](http://releases.k8s.io/release-1.0/docs/proposals/autoscaling.md).
## Overview
The resource usage of a serving application usually varies over time: sometimes the demand for the application rises,
and sometimes it drops.
In Kubernetes version 1.0, a user can only manually set the number of serving pods.
Our aim is to provide a mechanism for the automatic adjustment of the number of pods based on CPU utilization statistics
(a future version will allow autoscaling based on other resources/metrics).
The resource usage of a serving application usually varies over time: sometimes
the demand for the application rises, and sometimes it drops. In Kubernetes
version 1.0, a user can only manually set the number of serving pods. Our aim is
to provide a mechanism for the automatic adjustment of the number of pods based
on CPU utilization statistics (a future version will allow autoscaling based on
other resources/metrics).
## Scale Subresource
In Kubernetes version 1.1, we are introducing Scale subresource and implementing horizontal autoscaling of pods based on it.
Scale subresource is supported for replication controllers and deployments.
Scale subresource is a Virtual Resource (does not correspond to an object stored in etcd).
It is only present in the API as an interface that a controller (in this case the HorizontalPodAutoscaler) can use to dynamically scale
the number of replicas controlled by some other API object (currently ReplicationController and Deployment) and to learn the current number of replicas.
Scale is a subresource of the API object that it serves as the interface for.
The Scale subresource is useful because whenever we introduce another type we want to autoscale, we just need to implement the Scale subresource for it.
The wider discussion regarding Scale took place in [#1629](https://github.com/kubernetes/kubernetes/issues/1629).
In Kubernetes version 1.1, we are introducing Scale subresource and implementing
horizontal autoscaling of pods based on it. Scale subresource is supported for
replication controllers and deployments. Scale subresource is a Virtual Resource
(does not correspond to an object stored in etcd). It is only present in the API
as an interface that a controller (in this case the HorizontalPodAutoscaler) can
use to dynamically scale the number of replicas controlled by some other API
object (currently ReplicationController and Deployment) and to learn the current
number of replicas. Scale is a subresource of the API object that it serves as
the interface for. The Scale subresource is useful because whenever we introduce
another type we want to autoscale, we just need to implement the Scale
subresource for it. The wider discussion regarding Scale took place in issue
[#1629](https://github.com/kubernetes/kubernetes/issues/1629).
Scale subresource is in API for replication controller or deployment under the following paths:
Scale subresource is in API for replication controller or deployment under the
following paths:
`apis/extensions/v1beta1/replicationcontrollers/myrc/scale`
@ -99,14 +106,15 @@ type ScaleStatus struct {
}
```
Writing to `ScaleSpec.Replicas` resizes the replication controller/deployment associated with
the given Scale subresource.
`ScaleStatus.Replicas` reports how many pods are currently running in the replication controller/deployment,
and `ScaleStatus.Selector` returns selector for the pods.
Writing to `ScaleSpec.Replicas` resizes the replication controller/deployment
associated with the given Scale subresource. `ScaleStatus.Replicas` reports how
many pods are currently running in the replication controller/deployment, and
`ScaleStatus.Selector` returns selector for the pods.
## HorizontalPodAutoscaler Object
In Kubernetes version 1.1, we are introducing HorizontalPodAutoscaler object. It is accessible under:
In Kubernetes version 1.1, we are introducing HorizontalPodAutoscaler object. It
is accessible under:
`apis/extensions/v1beta1/horizontalpodautoscalers/myautoscaler`
@ -168,8 +176,9 @@ type HorizontalPodAutoscalerStatus struct {
```
`ScaleRef` is a reference to the Scale subresource.
`MinReplicas`, `MaxReplicas` and `CPUUtilization` define autoscaler configuration.
We are also introducing HorizontalPodAutoscalerList object to enable listing all autoscalers in a namespace:
`MinReplicas`, `MaxReplicas` and `CPUUtilization` define autoscaler
configuration. We are also introducing HorizontalPodAutoscalerList object to
enable listing all autoscalers in a namespace:
```go
// list of horizontal pod autoscaler objects.
@ -184,19 +193,22 @@ type HorizontalPodAutoscalerList struct {
## Autoscaling Algorithm
The autoscaler is implemented as a control loop. It periodically queries pods described by `Status.PodSelector` of Scale subresource, and collects their CPU utilization.
Then, it compares the arithmetic mean of the pods' CPU utilization with the target defined in `Spec.CPUUtilization`,
and adjust the replicas of the Scale if needed to match the target
(preserving condition: MinReplicas <= Replicas <= MaxReplicas).
The autoscaler is implemented as a control loop. It periodically queries pods
described by `Status.PodSelector` of Scale subresource, and collects their CPU
utilization. Then, it compares the arithmetic mean of the pods' CPU utilization
with the target defined in `Spec.CPUUtilization`, and adjusts the replicas of
the Scale if needed to match the target (preserving condition: MinReplicas <=
Replicas <= MaxReplicas).
The period of the autoscaler is controlled by `--horizontal-pod-autoscaler-sync-period` flag of controller manager.
The default value is 30 seconds.
The period of the autoscaler is controlled by the
`--horizontal-pod-autoscaler-sync-period` flag of controller manager. The
default value is 30 seconds.
CPU utilization is the recent CPU usage of a pod (average across the last 1 minute) divided by the CPU requested by the pod.
In Kubernetes version 1.1, CPU usage is taken directly from Heapster.
In future, there will be API on master for this purpose
(see [#11951](https://github.com/kubernetes/kubernetes/issues/11951)).
CPU utilization is the recent CPU usage of a pod (average across the last 1
minute) divided by the CPU requested by the pod. In Kubernetes version 1.1, CPU
usage is taken directly from Heapster. In future, there will be API on master
for this purpose (see issue [#11951](https://github.com/kubernetes/kubernetes/issues/11951)).
The target number of pods is calculated from the following formula:
@ -204,66 +216,76 @@ The target number of pods is calculated from the following formula:
TargetNumOfPods = ceil(sum(CurrentPodsCPUUtilization) / Target)
```
Starting and stopping pods may introduce noise to the metric (for instance, starting may temporarily increase CPU).
So, after each action, the autoscaler should wait some time for reliable data.
Scale-up can only happen if there was no rescaling within the last 3 minutes.
Scale-down will wait for 5 minutes from the last rescaling.
Moreover any scaling will only be made if: `avg(CurrentPodsConsumption) / Target` drops below 0.9 or increases above 1.1 (10% tolerance).
Such approach has two benefits:
Starting and stopping pods may introduce noise to the metric (for instance,
starting may temporarily increase CPU). So, after each action, the autoscaler
should wait some time for reliable data. Scale-up can only happen if there was
no rescaling within the last 3 minutes. Scale-down will wait for 5 minutes from
the last rescaling. Moreover any scaling will only be made if:
`avg(CurrentPodsConsumption) / Target` drops below 0.9 or increases above 1.1
(10% tolerance). Such approach has two benefits:
* Autoscaler works in a conservative way.
If new user load appears, it is important for us to rapidly increase the number of pods,
so that user requests will not be rejected.
Lowering the number of pods is not that urgent.
* Autoscaler works in a conservative way. If new user load appears, it is
important for us to rapidly increase the number of pods, so that user requests
will not be rejected. Lowering the number of pods is not that urgent.
* Autoscaler avoids thrashing, i.e.: prevents rapid execution of conflicting decision if the load is not stable.
* Autoscaler avoids thrashing, i.e.: prevents rapid execution of conflicting
decision if the load is not stable.
## Relative vs. absolute metrics
We chose values of the target metric to be relative (e.g. 90% of requested CPU resource) rather than absolute (e.g. 0.6 core) for the following reason.
If we choose absolute metric, user will need to guarantee that the target is lower than the request.
Otherwise, overloaded pods may not be able to consume more than the autoscaler's absolute target utilization,
thereby preventing the autoscaler from seeing high enough utilization to trigger it to scale up.
This may be especially troublesome when user changes requested resources for a pod
We chose values of the target metric to be relative (e.g. 90% of requested CPU
resource) rather than absolute (e.g. 0.6 core) for the following reason. If we
choose absolute metric, user will need to guarantee that the target is lower
than the request. Otherwise, overloaded pods may not be able to consume more
than the autoscaler's absolute target utilization, thereby preventing the
autoscaler from seeing high enough utilization to trigger it to scale up. This
may be especially troublesome when user changes requested resources for a pod
because they would need to also change the autoscaler utilization threshold.
Therefore, we decided to choose relative metric.
For user, it is enough to set it to a value smaller than 100%, and further changes of requested resources will not invalidate it.
Therefore, we decided to choose relative metric. For user, it is enough to set
it to a value smaller than 100%, and further changes of requested resources will
not invalidate it.
## Support in kubectl
To make manipulation of HorizontalPodAutoscaler object simpler, we added support for
creating/updating/deleting/listing of HorizontalPodAutoscaler to kubectl.
In addition, in future, we are planning to add kubectl support for the following use-cases:
* When creating a replication controller or deployment with `kubectl create [-f]`, there should be
a possibility to specify an additional autoscaler object.
(This should work out-of-the-box when creation of autoscaler is supported by kubectl as we may include
multiple objects in the same config file).
* *[future]* When running an image with `kubectl run`, there should be an additional option to create
an autoscaler for it.
* *[future]* We will add a new command `kubectl autoscale` that will allow for easy creation of an autoscaler object
for already existing replication controller/deployment.
To make manipulation of HorizontalPodAutoscaler object simpler, we added support
for creating/updating/deleting/listing of HorizontalPodAutoscaler to kubectl. In
addition, in future, we are planning to add kubectl support for the following
use-cases:
* When creating a replication controller or deployment with
`kubectl create [-f]`, there should be a possibility to specify an additional
autoscaler object. (This should work out-of-the-box when creation of autoscaler
is supported by kubectl as we may include multiple objects in the same config
file).
* *[future]* When running an image with `kubectl run`, there should be an
additional option to create an autoscaler for it.
* *[future]* We will add a new command `kubectl autoscale` that will allow for
easy creation of an autoscaler object for already existing replication
controller/deployment.
## Next steps
We list here some features that are not supported in Kubernetes version 1.1.
However, we want to keep them in mind, as they will most probably be needed in future.
However, we want to keep them in mind, as they will most probably be needed in
the future.
Our design is in general compatible with them.
* *[future]* **Autoscale pods based on metrics different than CPU** (e.g. memory, network traffic, qps).
This includes scaling based on a custom/application metric.
* *[future]* **Autoscale pods base on an aggregate metric.**
Autoscaler, instead of computing average for a target metric across pods, will use a single, external, metric (e.g. qps metric from load balancer).
The metric will be aggregated while the target will remain per-pod
(e.g. when observing 100 qps on load balancer while the target is 20 qps per pod, autoscaler will set the number of replicas to 5).
* *[future]* **Autoscale pods based on multiple metrics.**
If the target numbers of pods for different metrics are different, choose the largest target number of pods.
* *[future]* **Scale the number of pods starting from 0.**
All pods can be turned-off, and then turned-on when there is a demand for them.
When a request to service with no pods arrives, kube-proxy will generate an event for autoscaler
to create a new pod.
Discussed in [#3247](https://github.com/kubernetes/kubernetes/issues/3247).
* *[future]* **When scaling down, make more educated decision which pods to kill.**
E.g.: if two or more pods from the same replication controller are on the same node, kill one of them.
Discussed in [#4301](https://github.com/kubernetes/kubernetes/issues/4301).
* *[future]* **Autoscale pods based on metrics different than CPU** (e.g.
memory, network traffic, qps). This includes scaling based on a custom/application metric.
* *[future]* **Autoscale pods base on an aggregate metric.** Autoscaler,
instead of computing average for a target metric across pods, will use a single,
external, metric (e.g. qps metric from load balancer). The metric will be
aggregated while the target will remain per-pod (e.g. when observing 100 qps on
load balancer while the target is 20 qps per pod, autoscaler will set the number
of replicas to 5).
* *[future]* **Autoscale pods based on multiple metrics.** If the target numbers
of pods for different metrics are different, choose the largest target number of
pods.
* *[future]* **Scale the number of pods starting from 0.** All pods can be
turned-off, and then turned-on when there is a demand for them. When a request
to service with no pods arrives, kube-proxy will generate an event for
autoscaler to create a new pod. Discussed in issue [#3247](https://github.com/kubernetes/kubernetes/issues/3247).
* *[future]* **When scaling down, make more educated decision which pods to
kill.** E.g.: if two or more pods from the same replication controller are on
the same node, kill one of them. Discussed in issue [#4301](https://github.com/kubernetes/kubernetes/issues/4301).

108
docs/design/identifiers.md
View File

@ -34,95 +34,111 @@ Documentation for other releases can be found at
# Identifiers and Names in Kubernetes
A summarization of the goals and recommendations for identifiers in Kubernetes. Described in [GitHub issue #199](http://issue.k8s.io/199).
A summarization of the goals and recommendations for identifiers in Kubernetes.
Described in GitHub issue [#199](http://issue.k8s.io/199).
## Definitions
UID
: A non-empty, opaque, system-generated value guaranteed to be unique in time and space; intended to distinguish between historical occurrences of similar entities.
`UID`: A non-empty, opaque, system-generated value guaranteed to be unique in time
and space; intended to distinguish between historical occurrences of similar
entities.
Name
: A non-empty string guaranteed to be unique within a given scope at a particular time; used in resource URLs; provided by clients at creation time and encouraged to be human friendly; intended to facilitate creation idempotence and space-uniqueness of singleton objects, distinguish distinct entities, and reference particular entities across operations.
`Name`: A non-empty string guaranteed to be unique within a given scope at a
particular time; used in resource URLs; provided by clients at creation time and
encouraged to be human friendly; intended to facilitate creation idempotence and
space-uniqueness of singleton objects, distinguish distinct entities, and
reference particular entities across operations.
[rfc1035](http://www.ietf.org/rfc/rfc1035.txt)/[rfc1123](http://www.ietf.org/rfc/rfc1123.txt) label (DNS_LABEL)
: An alphanumeric (a-z, and 0-9) string, with a maximum length of 63 characters, with the '-' character allowed anywhere except the first or last character, suitable for use as a hostname or segment in a domain name
[rfc1035](http://www.ietf.org/rfc/rfc1035.txt)/[rfc1123](http://www.ietf.org/rfc/rfc1123.txt) `label` (DNS_LABEL):
An alphanumeric (a-z, and 0-9) string, with a maximum length of 63 characters,
with the '-' character allowed anywhere except the first or last character,
suitable for use as a hostname or segment in a domain name.
[rfc1035](http://www.ietf.org/rfc/rfc1035.txt)/[rfc1123](http://www.ietf.org/rfc/rfc1123.txt) subdomain (DNS_SUBDOMAIN)
: One or more lowercase rfc1035/rfc1123 labels separated by '.' with a maximum length of 253 characters
[rfc1035](http://www.ietf.org/rfc/rfc1035.txt)/[rfc1123](http://www.ietf.org/rfc/rfc1123.txt) `subdomain` (DNS_SUBDOMAIN):
One or more lowercase rfc1035/rfc1123 labels separated by '.' with a maximum
length of 253 characters.
[rfc4122](http://www.ietf.org/rfc/rfc4122.txt) universally unique identifier (UUID)
: A 128 bit generated value that is extremely unlikely to collide across time and space and requires no central coordination
[rfc4122](http://www.ietf.org/rfc/rfc4122.txt) `universally unique identifier` (UUID):
A 128 bit generated value that is extremely unlikely to collide across time and
space and requires no central coordination.
[rfc6335](https://tools.ietf.org/rfc/rfc6335.txt) port name (IANA_SVC_NAME)
: An alphanumeric (a-z, and 0-9) string, with a maximum length of 15 characters, with the '-' character allowed anywhere except the first or the last character or adjacent to another '-' character, it must contain at least a (a-z) character
[rfc6335](https://tools.ietf.org/rfc/rfc6335.txt) `port name` (IANA_SVC_NAME):
An alphanumeric (a-z, and 0-9) string, with a maximum length of 15 characters,
with the '-' character allowed anywhere except the first or the last character
or adjacent to another '-' character, it must contain at least a (a-z)
character.
## Objectives for names and UIDs
1. Uniquely identify (via a UID) an object across space and time
2. Uniquely name (via a name) an object across space
3. Provide human-friendly names in API operations and/or configuration files
4. Allow idempotent creation of API resources (#148) and enforcement of space-uniqueness of singleton objects
5. Allow DNS names to be automatically generated for some objects
1. Uniquely identify (via a UID) an object across space and time.
2. Uniquely name (via a name) an object across space.
3. Provide human-friendly names in API operations and/or configuration files.
4. Allow idempotent creation of API resources (#148) and enforcement of
space-uniqueness of singleton objects.
5. Allow DNS names to be automatically generated for some objects.
## General design
1. When an object is created via an API, a Name string (a DNS_SUBDOMAIN) must be specified. Name must be non-empty and unique within the apiserver. This enables idempotent and space-unique creation operations. Parts of the system (e.g. replication controller) may join strings (e.g. a base name and a random suffix) to create a unique Name. For situations where generating a name is impractical, some or all objects may support a param to auto-generate a name. Generating random names will defeat idempotency.
1. When an object is created via an API, a Name string (a DNS_SUBDOMAIN) must
be specified. Name must be non-empty and unique within the apiserver. This
enables idempotent and space-unique creation operations. Parts of the system
(e.g. replication controller) may join strings (e.g. a base name and a random
suffix) to create a unique Name. For situations where generating a name is
impractical, some or all objects may support a param to auto-generate a name.
Generating random names will defeat idempotency.
* Examples: "guestbook.user", "backend-x4eb1"
2. When an object is created via an API, a Namespace string (a DNS_SUBDOMAIN? format TBD via #1114) may be specified. Depending on the API receiver, namespaces might be validated (e.g. apiserver might ensure that the namespace actually exists). If a namespace is not specified, one will be assigned by the API receiver. This assignment policy might vary across API receivers (e.g. apiserver might have a default, kubelet might generate something semi-random).
2. When an object is created via an API, a Namespace string (a DNS_SUBDOMAIN?
format TBD via #1114) may be specified. Depending on the API receiver,
namespaces might be validated (e.g. apiserver might ensure that the namespace
actually exists). If a namespace is not specified, one will be assigned by the
API receiver. This assignment policy might vary across API receivers (e.g.
apiserver might have a default, kubelet might generate something semi-random).
* Example: "api.k8s.example.com"
3. Upon acceptance of an object via an API, the object is assigned a UID (a UUID). UID must be non-empty and unique across space and time.
3. Upon acceptance of an object via an API, the object is assigned a UID
(a UUID). UID must be non-empty and unique across space and time.
* Example: "01234567-89ab-cdef-0123-456789abcdef"
## Case study: Scheduling a pod
Pods can be placed onto a particular node in a number of ways. This case
study demonstrates how the above design can be applied to satisfy the
objectives.
Pods can be placed onto a particular node in a number of ways. This case study
demonstrates how the above design can be applied to satisfy the objectives.
### A pod scheduled by a user through the apiserver
1. A user submits a pod with Namespace="" and Name="guestbook" to the apiserver.
2. The apiserver validates the input.
1. A default Namespace is assigned.
2. The pod name must be space-unique within the Namespace.
3. Each container within the pod has a name which must be space-unique within the pod.
3. Each container within the pod has a name which must be space-unique within
the pod.
3. The pod is accepted.
1. A new UID is assigned.
4. The pod is bound to a node.
1. The kubelet on the node is passed the pod's UID, Namespace, and Name.
5. Kubelet validates the input.
6. Kubelet runs the pod.
1. Each container is started up with enough metadata to distinguish the pod from whence it came.
2. Each attempt to run a container is assigned a UID (a string) that is unique across time.
* This may correspond to Docker's container ID.
1. Each container is started up with enough metadata to distinguish the pod
from whence it came.
2. Each attempt to run a container is assigned a UID (a string) that is
unique across time. * This may correspond to Docker's container ID.
### A pod placed by a config file on the node
1. A config file is stored on the node, containing a pod with UID="", Namespace="", and Name="cadvisor".
1. A config file is stored on the node, containing a pod with UID="",
Namespace="", and Name="cadvisor".
2. Kubelet validates the input.
1. Since UID is not provided, kubelet generates one.
2. Since Namespace is not provided, kubelet generates one.
1. The generated namespace should be deterministic and cluster-unique for the source, such as a hash of the hostname and file path.
1. The generated namespace should be deterministic and cluster-unique for
the source, such as a hash of the hostname and file path.
* E.g. Namespace="file-f4231812554558a718a01ca942782d81"
3. Kubelet runs the pod.
1. Each container is started up with enough metadata to distinguish the pod from whence it came.
2. Each attempt to run a container is assigned a UID (a string) that is unique across time.
1. Each container is started up with enough metadata to distinguish the pod
from whence it came.
2. Each attempt to run a container is assigned a UID (a string) that is
unique across time.
1. This may correspond to Docker's container ID.

420
docs/design/indexed-job.md
View File

@ -53,64 +53,66 @@ a third way to run embarrassingly parallel programs, with a focus on
ease of use.
This new style of Job is called an *indexed job*, because each Pod of the Job
is specialized to work on a particular *index* from a fixed length array of work items.
is specialized to work on a particular *index* from a fixed length array of work
items.
## Background
The Kubernetes [Job](../../docs/user-guide/jobs.md) already supports
the embarrassingly parallel use case through *workqueue jobs*.
While [workqueue jobs](../../docs/user-guide/jobs.md#job-patterns)
are very flexible, they can be difficult to use.
They: (1) typically require running a message queue
or other database service, (2) typically require modifications
to existing binaries and images and (3) subtle race conditions
are easy to overlook.
While [workqueue jobs](../../docs/user-guide/jobs.md#job-patterns) are very
flexible, they can be difficult to use. They: (1) typically require running a
message queue or other database service, (2) typically require modifications
to existing binaries and images and (3) subtle race conditions are easy to
overlook.
Users also have another option for parallel jobs: creating [multiple Job objects
from a template](hdocs/design/indexed-job.md#job-patterns).
For small numbers of Jobs, this is a fine choice. Labels make it easy to view and
delete multiple Job objects at once. But, that approach also has its drawbacks:
(1) for large levels of parallelism (hundreds or thousands of pods) this approach
means that listing all jobs presents too much information, (2) users want a single
source of information about the success or failure of what the user views as a single
from a template](hdocs/design/indexed-job.md#job-patterns). For small numbers of
Jobs, this is a fine choice. Labels make it easy to view and delete multiple Job
objects at once. But, that approach also has its drawbacks: (1) for large levels
of parallelism (hundreds or thousands of pods) this approach means that listing
all jobs presents too much information, (2) users want a single source of
information about the success or failure of what the user views as a single
logical process.
Indexed job fills provides a third option with better ease-of-use for common use cases.
Indexed job fills provides a third option with better ease-of-use for common
use cases.
## Requirements
### User Requirements
- Users want an easy way to run a Pod to completion *for each* item within a
[work list](#example-use-cases).
[work list](#example-use-cases).
- Users want to run these pods in parallel for speed, but to vary the level of
parallelism as needed, independent of the number of work items.
parallelism as needed, independent of the number of work items.
- Users want to do this without requiring changes to existing images,
or source-to-image pipelines.
- Users want a single object that encompasses the lifetime of the parallel
program. Deleting it should delete all dependent objects. It should report
the status of the overall process. Users should be
able to wait for it to complete, and can refer to it from other resource types, such as
[ScheduledJob](https://github.com/kubernetes/kubernetes/pull/11980).
program. Deleting it should delete all dependent objects. It should report the
status of the overall process. Users should be able to wait for it to complete,
and can refer to it from other resource types, such as
[ScheduledJob](https://github.com/kubernetes/kubernetes/pull/11980).
### Example Use Cases
Here are several examples of *work lists*: lists of command lines that the
user wants to run, each line its own Pod. (Note that in practice, a work
list may not ever be written out in this form, but it exists in the mind of
the Job creator, and it is a useful way to talk about the the intent of the user when discussing alternatives for specifying Indexed Jobs).
Here are several examples of *work lists*: lists of command lines that the user
wants to run, each line its own Pod. (Note that in practice, a work list may not
ever be written out in this form, but it exists in the mind of the Job creator,
and it is a useful way to talk about the the intent of the user when discussing
alternatives for specifying Indexed Jobs).
Note that we will not have the user express their requirements in work list
form; it is just a format for presenting use cases. Subsequent discussion
will reference these work lists.
form; it is just a format for presenting use cases. Subsequent discussion will
reference these work lists.
#### Work List 1
Process several files with the same program
Process several files with the same program:
```
/usr/local/bin/process_file 12342.dat
@ -120,7 +122,7 @@ Process several files with the same program
#### Work List 2
Process a matrix (or image, etc) in rectangular blocks
Process a matrix (or image, etc) in rectangular blocks:
```
/usr/local/bin/process_matrix_block -start_row 0 -end_row 15 -start_col 0 --end_col 15
@ -131,7 +133,7 @@ Process a matrix (or image, etc) in rectangular blocks
#### Work List 3
Build a program at several different git commits
Build a program at several different git commits:
```
HASH=3cab5cb4a git checkout $HASH && make clean && make VERSION=$HASH
@ -141,7 +143,7 @@ HASH=a8b5e34c5 git checkout $HASH && make clean && make VERSION=$HASH
#### Work List 4
Render several frames of a movie.
Render several frames of a movie:
```
./blender /vol1/mymodel.blend -o /vol2/frame_#### -f 1
@ -151,7 +153,8 @@ Render several frames of a movie.
#### Work List 5
Render several blocks of frames. (Render blocks to avoid Pod startup overhead for every frame)
Render several blocks of frames (Render blocks to avoid Pod startup overhead for
every frame):
```
./blender /vol1/mymodel.blend -o /vol2/frame_#### --frame-start 1 --frame-end 100
@ -167,57 +170,59 @@ Given a work list, like in the [work list examples](#work-list-examples),
the information from the work list needs to get into each Pod of the Job.
Users will typically not want to create a new image for each job they
run. They will want to use existing images. So, the image is not the place
run. They will want to use existing images. So, the image is not the place
for the work list.
A work list can be stored on networked storage, and mounted by pods of the job.
Also, as a shortcut, for small worklists, it can be included in an annotation on the Job object,
which is then exposed as a volume in the pod via the downward API.
Also, as a shortcut, for small worklists, it can be included in an annotation on
the Job object, which is then exposed as a volume in the pod via the downward
API.
### What Varies Between Pods of a Job
Pods need to differ in some way to do something different. (They do not
differ in the work-queue style of Job, but that style has ease-of-use issues).
Pods need to differ in some way to do something different. (They do not differ
in the work-queue style of Job, but that style has ease-of-use issues).
A general approach would be to allow pods to differ from each other in arbitrary ways.
For example, the Job object could have a list of PodSpecs to run.
However, this is so general that it provides little value. It would:
A general approach would be to allow pods to differ from each other in arbitrary
ways. For example, the Job object could have a list of PodSpecs to run.
However, this is so general that it provides little value. It would:
- make the Job Spec very verbose, especially for jobs with thousands of work items
- make the Job Spec very verbose, especially for jobs with thousands of work
items
- Job becomes such a vague concept that it is hard to explain to users
- in practice, we do not see cases where many pods which differ across many fields of their
specs, and need to run as a group, with no ordering constraints.
- in practice, we do not see cases where many pods which differ across many
fields of their specs, and need to run as a group, with no ordering constraints.
- CLIs and UIs need to support more options for creating Job
- it is useful for monitoring and accounting databases want to aggregate data for pods
with the same controller. However, pods with very different Specs may not make sense
to aggregate.
- profiling, debugging, accounting, auditing and monitoring tools cannot assume common
images/files, behaviors, provenance and so on between Pods of a Job.
- it is useful for monitoring and accounting databases want to aggregate data
for pods with the same controller. However, pods with very different Specs may
not make sense to aggregate.
- profiling, debugging, accounting, auditing and monitoring tools cannot assume
common images/files, behaviors, provenance and so on between Pods of a Job.
Also, variety has another cost. Pods which differ in ways that affect scheduling
(node constraints, resource requirements, labels) prevent the scheduler
from treating them as fungible, which is an important optimization for the scheduler.
Also, variety has another cost. Pods which differ in ways that affect scheduling
(node constraints, resource requirements, labels) prevent the scheduler from
treating them as fungible, which is an important optimization for the scheduler.
Therefore, we will not allow Pods from the same Job to differ arbitrarily
(anyway, users can use multiple Job objects for that case). We will try to
allow as little as possible to differ between pods of the same Job, while
still allowing users to express common parallel patterns easily.
For users who need to run jobs which differ in other ways, they can create multiple
Jobs, and manage them as a group using labels.
allow as little as possible to differ between pods of the same Job, while still
allowing users to express common parallel patterns easily. For users who need to
run jobs which differ in other ways, they can create multiple Jobs, and manage
them as a group using labels.
From the above work lists, we see a need for Pods which differ in their command
lines, and in their environment variables. These work lists do not require the
pods to differ in other ways.
Experience in a [similar systems](http://static.googleusercontent.com/media/research.google.com/en//pubs/archive/43438.pdf) has shown this model to be applicable
to a very broad range of problems, despite this restriction.
Therefore we to allow pods in the same Job to differ **only** in the following aspects:
Experience in [similar systems](http://static.googleusercontent.com/media/research.google.com/en//pubs/archive/43438.pdf)
has shown this model to be applicable to a very broad range of problems, despite
this restriction.
Therefore we to allow pods in the same Job to differ **only** in the following
aspects:
- command line
- environment variables
### Composition of existing images
The docker image that is used in a job may not be maintained by the person
@ -230,9 +235,9 @@ This needs more thought.
### Running Ad-Hoc Jobs using kubectl
A user should be able to easily start an Indexed Job using `kubectl`.
For example to run [work list 1](#work-list-1), a user should be able
to type something simple like:
A user should be able to easily start an Indexed Job using `kubectl`. For
example to run [work list 1](#work-list-1), a user should be able to type
something simple like:
```
kubectl run process-files --image=myfileprocessor \
@ -246,13 +251,16 @@ In the above example:
- `--restart=OnFailure` implies creating a job instead of replicationController.
- Each pods command line is `/usr/local/bin/process_file $F`.
- `--per-completion-env=` implies the jobs `.spec.completions` is set to the length of the argument array (3 in the example).
- `--per-completion-env=F=<values>` causes env var with `F` to be available in the environment when the command line is evaluated.
- `--per-completion-env=` implies the jobs `.spec.completions` is set to the
length of the argument array (3 in the example).
- `--per-completion-env=F=<values>` causes env var with `F` to be available in
the environment when the command line is evaluated.
How exactly this happens is discussed later in the doc: this is a sketch of the user experience.
How exactly this happens is discussed later in the doc: this is a sketch of the
user experience.
In practice, the list of files might be much longer and stored in a file
on the users local host, like:
In practice, the list of files might be much longer and stored in a file on the
users local host, like:
```
$ cat files-to-process.txt
@ -266,16 +274,27 @@ So, the user could specify instead: `--per-completion-env=F="$(cat files-to-proc
However, `kubectl` should also support a format like:
`--per-completion-env=F=@files-to-process.txt`.
That allows `kubectl` to parse the file, point out any syntax errors, and would not run up against command line length limits (2MB is common, as low as 4kB is POSIX compliant).
That allows `kubectl` to parse the file, point out any syntax errors, and would
not run up against command line length limits (2MB is common, as low as 4kB is
POSIX compliant).
One case we do not try to handle is where the file of work is stored on a cloud filesystem, and not accessible from the users local host. Then we cannot easily use indexed job, because we do not know the number of completions. The user needs to copy the file locally first or use the Work-Queue style of Job (already supported).
One case we do not try to handle is where the file of work is stored on a cloud
filesystem, and not accessible from the users local host. Then we cannot easily
use indexed job, because we do not know the number of completions. The user
needs to copy the file locally first or use the Work-Queue style of Job (already
supported).
Another case we do not try to handle is where the input file does not exist yet because this Job is to be run at a future time, or depends on another job. The workflow and scheduled job proposal need to consider this case. For that case, you could use an indexed job which runs a program which shards the input file (map-reduce-style).
Another case we do not try to handle is where the input file does not exist yet
because this Job is to be run at a future time, or depends on another job. The
workflow and scheduled job proposal need to consider this case. For that case,
you could use an indexed job which runs a program which shards the input file
(map-reduce-style).
#### Multiple parameters
The user may also have multiple parameters, like in [work list 2](#work-list-2).
One way is to just list all the command lines already expanded, one per line, in a file, like this:
One way is to just list all the command lines already expanded, one per line, in
a file, like this:
```
$ cat matrix-commandlines.txt
@ -295,10 +314,12 @@ kubectl run process-matrix --image=my/matrix \
'eval "$COMMAND_LINE"'
```
However, this may have some subtleties with shell escaping. Also, it depends on the user
knowing all the correct arguments to the docker image being used (more on this later).
However, this may have some subtleties with shell escaping. Also, it depends on
the user knowing all the correct arguments to the docker image being used (more
on this later).
Instead, kubectl should support multiple instances of the `--per-completion-env` flag. For example, to implement work list 2, a user could do:
Instead, kubectl should support multiple instances of the `--per-completion-env`
flag. For example, to implement work list 2, a user could do:
```
kubectl run process-matrix --image=my/matrix \
@ -313,8 +334,8 @@ kubectl run process-matrix --image=my/matrix \
### Composition With Workflows and ScheduledJob
A user should be able to create a job (Indexed or not) which runs at a specific time(s).
For example:
A user should be able to create a job (Indexed or not) which runs at a specific
time(s). For example:
```
$ kubectl run process-files --image=myfileprocessor \
@ -326,12 +347,16 @@ $ kubectl run process-files --image=myfileprocessor \
created "scheduledJob/process-files-37dt3"
```
Kubectl should build the same JobSpec, and then put it into a ScheduledJob (#11980) and create that.
Kubectl should build the same JobSpec, and then put it into a ScheduledJob
(#11980) and create that.
For [workflow type jobs](../../docs/user-guide/jobs.md#job-patterns), creating a complete workflow from a single command line would be messy, because of the need to specify all the arguments multiple times.
For [workflow type jobs](../../docs/user-guide/jobs.md#job-patterns), creating a
complete workflow from a single command line would be messy, because of the need
to specify all the arguments multiple times.
For that use case, the user could create a workflow message by hand.
Or the user could create a job template, and then make a workflow from the templates, perhaps like this:
For that use case, the user could create a workflow message by hand. Or the user
could create a job template, and then make a workflow from the templates,
perhaps like this:
```
$ kubectl run process-files --image=myfileprocessor \
@ -357,17 +382,17 @@ created "workflow/process-and-merge"
### Completion Indexes
A JobSpec specifies the number of times a pod needs to complete successfully,
through the `job.Spec.Completions` field. The number of completions
will be equal to the number of work items in the work list.
through the `job.Spec.Completions` field. The number of completions will be
equal to the number of work items in the work list.
Each pod that the job controller creates is intended to complete one work item
from the work list. Since a pod may fail, several pods may, serially,
attempt to complete the same index. Therefore, we call it a
a *completion index* (or just *index*), but not a *pod index*.
from the work list. Since a pod may fail, several pods may, serially, attempt to
complete the same index. Therefore, we call it a a *completion index* (or just
*index*), but not a *pod index*.
For each completion index, in the range 1 to `.job.Spec.Completions`,
the job controller will create a pod with that index, and keep creating them
on failure, until each index is completed.
For each completion index, in the range 1 to `.job.Spec.Completions`, the job
controller will create a pod with that index, and keep creating them on failure,
until each index is completed.
An dense integer index, rather than a sparse string index (e.g. using just
`metadata.generate-name`) makes it easy to use the index to lookup parameters
@ -375,9 +400,9 @@ in, for example, an array in shared storage.
### Pod Identity and Template Substitution in Job Controller
The JobSpec contains a single pod template. When the job controller creates a particular
pod, it copies the pod template and modifies it in some way to make that pod distinctive.
Whatever is distinctive about that pod is its *identity*.
The JobSpec contains a single pod template. When the job controller creates a
particular pod, it copies the pod template and modifies it in some way to make
that pod distinctive. Whatever is distinctive about that pod is its *identity*.
We consider several options.
@ -387,45 +412,46 @@ The job controller substitutes only the *completion index* of the pod into the
pod template when creating it. The JSON it POSTs differs only in a single
fields.
We would put the completion index as a stringified integer, into an
annotation of the pod. The user can extract it from the annotation
into an env var via the downward API, or put it in a file via a Downward
API volume, and parse it himself.
We would put the completion index as a stringified integer, into an annotation
of the pod. The user can extract it from the annotation into an env var via the
downward API, or put it in a file via a Downward API volume, and parse it
himself.
Once it is an environment variable in the pod (say `$INDEX`),
then one of two things can happen.
Once it is an environment variable in the pod (say `$INDEX`), then one of two
things can happen.
First, the main program can know how to map from an integer index to what it
needs to do.
For example, from Work List 4 above:
needs to do. For example, from Work List 4 above:
```
./blender /vol1/mymodel.blend -o /vol2/frame_#### -f $INDEX
```
Second, a shell script can be prepended to the original command line which maps the
index to one or more string parameters. For example, to implement Work List 5 above,
you could do:
Second, a shell script can be prepended to the original command line which maps
the index to one or more string parameters. For example, to implement Work List
5 above, you could do:
```
/vol0/setupenv.sh && ./blender /vol1/mymodel.blend -o /vol2/frame_#### --frame-start $START_FRAME --frame-end $END_FRAME
```
In the above example, `/vol0/setupenv.sh` is a shell script that reads `$INDEX` and exports `$START_FRAME` and `$END_FRAME`.
In the above example, `/vol0/setupenv.sh` is a shell script that reads `$INDEX`
and exports `$START_FRAME` and `$END_FRAME`.
The shell could be part of the image, but more usefully, it could be generated by a program and stuffed in an annotation
or a configMap, and from there added to a volume.
The shell could be part of the image, but more usefully, it could be generated
by a program and stuffed in an annotation or a configMap, and from there added
to a volume.
The first approach may require the user
to modify an existing image (see next section) to be able to accept an `$INDEX` env var or argument.
The second approach requires that the image have a shell. We think that together these two options
cover a wide range of use cases (though not all).
The first approach may require the user to modify an existing image (see next
section) to be able to accept an `$INDEX` env var or argument. The second
approach requires that the image have a shell. We think that together these two
options cover a wide range of use cases (though not all).
#### Multiple Substitution
In this option, the JobSpec is extended to include a list of values to substitute,
and which fields to substitute them into. For example, a worklist like this:
In this option, the JobSpec is extended to include a list of values to
substitute, and which fields to substitute them into. For example, a worklist
like this:
```
FRUIT_COLOR=green process-fruit -a -b -c -f apple.txt --remove-seeds
@ -433,7 +459,7 @@ FRUIT_COLOR=yellow process-fruit -a -b -c -f banana.txt
FRUIT_COLOR=red process-fruit -a -b -c -f cherry.txt --remove-pit
```
Can be broken down into a template like this, with three parameters
Can be broken down into a template like this, with three parameters:
```
<custom env var 1>; process-fruit -a -b -c <custom arg 1> <custom arg 1>
@ -447,9 +473,8 @@ and a list of parameter tuples, like this:
("FRUIT_COLOR=red", "-f cherry.txt", "--remove-pit")
```
The JobSpec can be extended to hold a list of parameter tuples (which
are more easily expressed as a list of lists of individual parameters).
For example:
The JobSpec can be extended to hold a list of parameter tuples (which are more
easily expressed as a list of lists of individual parameters). For example:
```
apiVersion: extensions/v1beta1
@ -477,42 +502,46 @@ spec:
- "red"
```
However, just providing custom env vars, and not arguments, is sufficient
for many use cases: parameter can be put into env vars, and then
substituted on the command line.
However, just providing custom env vars, and not arguments, is sufficient for
many use cases: parameter can be put into env vars, and then substituted on the
command line.
#### Comparison
The multiple substitution approach:
- keeps the *per completion parameters* in the JobSpec.
- Drawback: makes the job spec large for job with thousands of completions. (But for very large jobs, the work-queue style or another type of controller, such as map-reduce or spark, may be a better fit.)
- Drawback: is a form of server-side templating, which we want in Kubernetes but have not fully designed
(see the [PetSets proposal](https://github.com/kubernetes/kubernetes/pull/18016/files?short_path=61f4179#diff-61f41798f4bced6e42e45731c1494cee)).
- Drawback: makes the job spec large for job with thousands of completions. (But
for very large jobs, the work-queue style or another type of controller, such as
map-reduce or spark, may be a better fit.)
- Drawback: is a form of server-side templating, which we want in Kubernetes but
have not fully designed (see the [PetSets proposal](https://github.com/kubernetes/kubernetes/pull/18016/files?short_path=61f4179#diff-61f41798f4bced6e42e45731c1494cee)).
The index-only approach:
- requires that the user keep the *per completion parameters* in a separate storage, such as a configData or networked storage.
- makes no changes to the JobSpec.
- Drawback: while in separate storage, they could be mutatated, which would have unexpected effects
- Requires that the user keep the *per completion parameters* in a separate
storage, such as a configData or networked storage.
- Makes no changes to the JobSpec.
- Drawback: while in separate storage, they could be mutatated, which would have
unexpected effects.
- Drawback: Logic for using index to lookup parameters needs to be in the Pod.
- Drawback: CLIs and UIs are limited to using the "index" as the identity of a pod
from a job. They cannot easily say, for example `repeated failures on the pod processing banana.txt`.
- Drawback: CLIs and UIs are limited to using the "index" as the identity of a
pod from a job. They cannot easily say, for example `repeated failures on the
pod processing banana.txt`.
Index-only approach relies on at least one of the following being true:
1. image containing a shell and certain shell commands (not all images have this)
1. use directly consumes the index from annoations (file or env var) and expands to specific behavior in the main program.
1. Image containing a shell and certain shell commands (not all images have
this).
1. Use directly consumes the index from annotations (file or env var) and
expands to specific behavior in the main program.
Also Using the index-only approach from
non-kubectl clients requires that they mimic the script-generation step,
or only use the second style.
Also Using the index-only approach from non-kubectl clients requires that they
mimic the script-generation step, or only use the second style.
#### Decision
It is decided to implement the Index-only approach now. Once the server-side
It is decided to implement the Index-only approach now. Once the server-side
templating design is complete for Kubernetes, and we have feedback from users,
we can consider if Multiple Substitution.
@ -523,43 +552,42 @@ we can consider if Multiple Substitution.
No changes are made to the JobSpec.
The JobStatus is also not changed.
The user can gauge the progress of the job by the `.status.succeeded` count.
The JobStatus is also not changed. The user can gauge the progress of the job by
the `.status.succeeded` count.
#### Job Spec Compatilibity
A job spec written before this change will work exactly the same
as before with the new controller.
The Pods it creates will have the same environment as before.
They will have a new annotation, but pod are expected to tolerate
A job spec written before this change will work exactly the same as before with
the new controller. The Pods it creates will have the same environment as
before. They will have a new annotation, but pod are expected to tolerate
unfamiliar annotations.
However, if the job controller version is reverted, to a version before this change,
the jobs whose pod specs depend on the the new annotation will fail. This is
okay for a Beta resource.
However, if the job controller version is reverted, to a version before this
change, the jobs whose pod specs depend on the the new annotation will fail.
This is okay for a Beta resource.
#### Job Controller Changes
The Job controller will maintain for each Job a data structed which
indicates the status of each completion index. We call this the
*scoreboard* for short. It is an array of length `.spec.completions`.
indicates the status of each completion index. We call this the
*scoreboard* for short. It is an array of length `.spec.completions`.
Elements of the array are `enum` type with possible values including
`complete`, `running`, and `notStarted`.
The scoreboard is stored in Job Controller
memory for efficiency. In either case, the Status can be reconstructed from
watching pods of the job (such as on a controller manager restart).
The index of the pods can be extracted from the pod annotation.
The scoreboard is stored in Job Controller memory for efficiency. In either
case, the Status can be reconstructed from watching pods of the job (such as on
a controller manager restart). The index of the pods can be extracted from the
pod annotation.
When Job controller sees that the number of running pods is less than the desired
parallelism of the job, it finds the first index in the scoreboard with value
`notRunning`. It creates a pod with this creation index.
When Job controller sees that the number of running pods is less than the
desired parallelism of the job, it finds the first index in the scoreboard with
value `notRunning`. It creates a pod with this creation index.
When it creates a pod with creation index `i`, it makes a copy
of the `.spec.template`, and sets
`.spec.template.metadata.annotations.[kubernetes.io/job/completion-index]`
to `i`. It does this in both the index-only and multiple-substitutions options.
When it creates a pod with creation index `i`, it makes a copy of the
`.spec.template`, and sets
`.spec.template.metadata.annotations.[kubernetes.io/job/completion-index]` to
`i`. It does this in both the index-only and multiple-substitutions options.
Then it creates the pod.
@ -571,8 +599,8 @@ When all entries in the scoreboard are `complete`, then the job is complete.
#### Downward API Changes
The downward API is changed to support extracting specific key names
into a single environment variable. So, the following would be supported:
The downward API is changed to support extracting specific key names into a
single environment variable. So, the following would be supported:
```
kind: Pod
@ -589,15 +617,16 @@ spec:
This requires kubelet changes.
Users who fail to upgrade their kubelets at the same time as they upgrade their controller
manager will see a failure for pods to run when they are created by the controller.
The Kubelet will send an event about failure to create the pod.
Users who fail to upgrade their kubelets at the same time as they upgrade their
controller manager will see a failure for pods to run when they are created by
the controller. The Kubelet will send an event about failure to create the pod.
The `kubectl describe job` will show many failed pods.
#### Kubectl Interface Changes
The `--completions` and `--completion-index-var-name` flags are added to kubectl.
The `--completions` and `--completion-index-var-name` flags are added to
kubectl.
For example, this command:
@ -621,8 +650,8 @@ Kubectl would create the following pod:
Kubectl will also support the `--per-completion-env` flag, as described previously.
For example, this command:
Kubectl will also support the `--per-completion-env` flag, as described
previously. For example, this command:
```
kubectl run say-fruit --image=busybox \
@ -655,7 +684,7 @@ kubectl run say-fruit --image=busybox \
sh -c 'echo "Have a nice $COLOR $FRUIT" && sleep 5'
```
will all run 3 pods in parallel. Index 0 pod will log:
will all run 3 pods in parallel. Index 0 pod will log:
```
Have a nice grenn apple
@ -666,16 +695,20 @@ and so on.
Notes:
- `--per-completion-env=` is of form `KEY=VALUES` where `VALUES` is either a quoted
space separated list or `@` and the name of a text file containing a list.
- `--per-completion-env=` can be specified several times, but all must have the same
length list
- `--per-completion-env=` is of form `KEY=VALUES` where `VALUES` is either a
quoted space separated list or `@` and the name of a text file containing a
list.
- `--per-completion-env=` can be specified several times, but all must have the
same length list.
- `--completions=N` with `N` equal to list length is implied.
- The flag `--completions=3` sets `job.spec.completions=3`.
- The flag `--completion-index-var-name=I` causes an env var to be created named I in each pod, with the index in it.
- The flag `--restart=OnFailure` is implied by `--completions` or any job-specific arguments. The user can also specify
`--restart=Never` if they desire but may not specify `--restart=Always` with job-related flags.
- Setting any of these flags in turn tells kubectl to create a Job, not a replicationController.
- The flag `--completion-index-var-name=I` causes an env var to be created named
I in each pod, with the index in it.
- The flag `--restart=OnFailure` is implied by `--completions` or any
job-specific arguments. The user can also specify `--restart=Never` if they
desire but may not specify `--restart=Always` with job-related flags.
- Setting any of these flags in turn tells kubectl to create a Job, not a
replicationController.
#### How Kubectl Creates Job Specs.
@ -850,14 +883,17 @@ configData/secret, and prevent the case where someone changes the
configData mid-job, and breaks things in a hard-to-debug way.
## Interactions with other features
#### Supporting Work Queue Jobs too
For Work Queue Jobs, completions has no meaning. Parallelism should be allowed to be greater than it, and pods have no identity. So, the job controller should not create a scoreboard in the JobStatus, just a count. Therefore, we need to add one of the following to JobSpec:
For Work Queue Jobs, completions has no meaning. Parallelism should be allowed
to be greater than it, and pods have no identity. So, the job controller should
not create a scoreboard in the JobStatus, just a count. Therefore, we need to
add one of the following to JobSpec:
- allow unset `.spec.completions` to indicate no scoreboard, and no index for tasks (identical tasks)
- allow unset `.spec.completions` to indicate no scoreboard, and no index for
tasks (identical tasks).
- allow `.spec.completions=-1` to indicate the same.
- add `.spec.indexed` to job to indicate need for scoreboard.
@ -866,33 +902,31 @@ For Work Queue Jobs, completions has no meaning. Parallelism should be allowed
Since pods of the same job will not be created with different resources,
a vertical autoscaler will need to:
- if it has index-specific initial resource suggestions, suggest those at admission
time; it will need to understand indexes.
- mutate resource requests on already created pods based on usage trend or previous container failures
- if it has index-specific initial resource suggestions, suggest those at
admission time; it will need to understand indexes.
- mutate resource requests on already created pods based on usage trend or
previous container failures.
- modify the job template, affecting all indexes.
#### Comparison to PetSets
The *Index substitution-only* option corresponds roughly to PetSet Proposal 1b.
The `perCompletionArgs` approach is similar to PetSet Proposal 1e, but more restrictive and thus less verbose.
The `perCompletionArgs` approach is similar to PetSet Proposal 1e, but more
restrictive and thus less verbose.
It would be easier for users if Indexed Job and PetSet are similar where possible.
However, PetSet differs in several key respects:
It would be easier for users if Indexed Job and PetSet are similar where
possible. However, PetSet differs in several key respects:
- PetSet is for ones to tens of instances. Indexed job should work with tens of
thousands of instances.
- When you have few instances, you may want to given them pet names. When you have many
instances, you that many instances, integer indexes make more sense.
thousands of instances.
- When you have few instances, you may want to given them pet names. When you
have many instances, you that many instances, integer indexes make more sense.
- When you have thousands of instances, storing the work-list in the JobSpec
is verbose. For PetSet, this is less of a problem.
is verbose. For PetSet, this is less of a problem.
- PetSets (apparently) need to differ in more fields than indexed Jobs.
This differs from PetSet in that PetSet uses names and not indexes.
PetSet is intended to support ones to tens of things.
This differs from PetSet in that PetSet uses names and not indexes. PetSet is
intended to support ones to tens of things.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

68
docs/design/metadata-policy.md
View File

@ -38,21 +38,24 @@ Documentation for other releases can be found at
This document describes a new API resource, `MetadataPolicy`, that configures an
admission controller to take one or more actions based on an object's metadata.
Initially the metadata fields that the predicates can examine are labels and annotations,
and the actions are to add one or more labels and/or annotations, or to reject creation/update
of the object. In the future other actions might be supported, such as applying an initializer.
Initially the metadata fields that the predicates can examine are labels and
annotations, and the actions are to add one or more labels and/or annotations,
or to reject creation/update of the object. In the future other actions might be
supported, such as applying an initializer.
The first use of `MetadataPolicy` will be to decide which scheduler should schedule a pod
in a [multi-scheduler](../proposals/multiple-schedulers.md) Kubernetes system. In particular, the
policy will add the scheduler name annotation to a pod based on an annotation that
is already on the pod that indicates the QoS of the pod.
(That annotation was presumably set by a simpler admission controller that
uses code, rather than configuration, to map the resource requests and limits of a pod
to QoS, and attaches the corresponding annotation.)
The first use of `MetadataPolicy` will be to decide which scheduler should
schedule a pod in a [multi-scheduler](../proposals/multiple-schedulers.md)
Kubernetes system. In particular, the policy will add the scheduler name
annotation to a pod based on an annotation that is already on the pod that
indicates the QoS of the pod. (That annotation was presumably set by a simpler
admission controller that uses code, rather than configuration, to map the
resource requests and limits of a pod to QoS, and attaches the corresponding
annotation.)
We anticipate a number of other uses for `MetadataPolicy`, such as defaulting for
labels and annotations, prohibiting/requiring particular labels or annotations, or
choosing a scheduling policy within a scheduler. We do not discuss them in this doc.
We anticipate a number of other uses for `MetadataPolicy`, such as defaulting
for labels and annotations, prohibiting/requiring particular labels or
annotations, or choosing a scheduling policy within a scheduler. We do not
discuss them in this doc.
## API
@ -126,7 +129,8 @@ type MetadataPolicyList struct {
## Implementation plan
1. Create `MetadataPolicy` API resource
1. Create admission controller that implements policies defined in `MetadataPolicy`
1. Create admission controller that implements policies defined in
`MetadataPolicy`
1. Create admission controller that sets annotation
`scheduler.alpha.kubernetes.io/qos: <QoS>`
(where `QOS` is one of `Guaranteed, Burstable, BestEffort`)
@ -134,30 +138,32 @@ based on pod's resource request and limit.
## Future work
Longer-term we will have QoS be set on create and update by the registry, similar to `Pending` phase today,
instead of having an admission controller (that runs before the one that takes `MetadataPolicy` as input)
do it.
Longer-term we will have QoS be set on create and update by the registry,
similar to `Pending` phase today, instead of having an admission controller
(that runs before the one that takes `MetadataPolicy` as input) do it.
We plan to eventually move from having an admission controller
set the scheduler name as a pod annotation, to using the initializer concept. In particular, the
scheduler will be an initializer, and the admission controller that decides which scheduler to use
will add the scheduler's name to the list of initializers for the pod (presumably the scheduler
will be the last initializer to run on each pod).
The admission controller would still be configured using the `MetadataPolicy` described here, only the
mechanism the admission controller uses to record its decision of which scheduler to use would change.
We plan to eventually move from having an admission controller set the scheduler
name as a pod annotation, to using the initializer concept. In particular, the
scheduler will be an initializer, and the admission controller that decides
which scheduler to use will add the scheduler's name to the list of initializers
for the pod (presumably the scheduler will be the last initializer to run on
each pod). The admission controller would still be configured using the
`MetadataPolicy` described here, only the mechanism the admission controller
uses to record its decision of which scheduler to use would change.
## Related issues
The main issue for multiple schedulers is #11793. There was also a lot of discussion
in PRs #17197 and #17865.
The main issue for multiple schedulers is #11793. There was also a lot of
discussion in PRs #17197 and #17865.
We could use the approach described here to choose a scheduling
policy within a single scheduler, as opposed to choosing a scheduler, a desire mentioned in #9920.
Issue #17097 describes a scenario unrelated to scheduler-choosing where `MetadataPolicy` could be used.
Issue #17324 proposes to create a generalized API for matching
"claims" to "service classes"; matching a pod to a scheduler would be one use for such an API.
We could use the approach described here to choose a scheduling policy within a
single scheduler, as opposed to choosing a scheduler, a desire mentioned in
# 9920. Issue #17097 describes a scenario unrelated to scheduler-choosing where
`MetadataPolicy` could be used. Issue #17324 proposes to create a generalized
API for matching "claims" to "service classes"; matching a pod to a scheduler
would be one use for such an API.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

180
docs/design/namespaces.md
View File

@ -41,9 +41,11 @@ a logically named group.
## Motivation
A single cluster should be able to satisfy the needs of multiple user communities.
A single cluster should be able to satisfy the needs of multiple user
communities.
Each user community wants to be able to work in isolation from other communities.
Each user community wants to be able to work in isolation from other
communities.
Each user community has its own:
@ -61,13 +63,16 @@ The Namespace provides a unique scope for:
## Use cases
1. As a cluster operator, I want to support multiple user communities on a single cluster.
2. As a cluster operator, I want to delegate authority to partitions of the cluster to trusted users
in those communities.
3. As a cluster operator, I want to limit the amount of resources each community can consume in order
to limit the impact to other communities using the cluster.
4. As a cluster user, I want to interact with resources that are pertinent to my user community in
isolation of what other user communities are doing on the cluster.
1. As a cluster operator, I want to support multiple user communities on a
single cluster.
2. As a cluster operator, I want to delegate authority to partitions of the
cluster to trusted users in those communities.
3. As a cluster operator, I want to limit the amount of resources each
community can consume in order to limit the impact to other communities using
the cluster.
4. As a cluster user, I want to interact with resources that are pertinent to
my user community in isolation of what other user communities are doing on the
cluster.
## Design
@ -91,20 +96,26 @@ A *Namespace* must exist prior to associating content with it.
A *Namespace* must not be deleted if there is content associated with it.
To associate a resource with a *Namespace* the following conditions must be satisfied:
To associate a resource with a *Namespace* the following conditions must be
satisfied:
1. The resource's *Kind* must be registered as having *RESTScopeNamespace* with the server
2. The resource's *TypeMeta.Namespace* field must have a value that references an existing *Namespace*
1. The resource's *Kind* must be registered as having *RESTScopeNamespace* with
the server
2. The resource's *TypeMeta.Namespace* field must have a value that references
an existing *Namespace*
The *Name* of a resource associated with a *Namespace* is unique to that *Kind* in that *Namespace*.
The *Name* of a resource associated with a *Namespace* is unique to that *Kind*
in that *Namespace*.
It is intended to be used in resource URLs; provided by clients at creation time, and encouraged to be
human friendly; intended to facilitate idempotent creation, space-uniqueness of singleton objects,
distinguish distinct entities, and reference particular entities across operations.
It is intended to be used in resource URLs; provided by clients at creation
time, and encouraged to be human friendly; intended to facilitate idempotent
creation, space-uniqueness of singleton objects, distinguish distinct entities,
and reference particular entities across operations.
### Authorization
A *Namespace* provides an authorization scope for accessing content associated with the *Namespace*.
A *Namespace* provides an authorization scope for accessing content associated
with the *Namespace*.
See [Authorization plugins](../admin/authorization.md)
@ -112,19 +123,21 @@ See [Authorization plugins](../admin/authorization.md)
A *Namespace* provides a scope to limit resource consumption.
A *LimitRange* defines min/max constraints on the amount of resources a single entity can consume in
a *Namespace*.
A *LimitRange* defines min/max constraints on the amount of resources a single
entity can consume in a *Namespace*.
See [Admission control: Limit Range](admission_control_limit_range.md)
A *ResourceQuota* tracks aggregate usage of resources in the *Namespace* and allows cluster operators
to define *Hard* resource usage limits that a *Namespace* may consume.
A *ResourceQuota* tracks aggregate usage of resources in the *Namespace* and
allows cluster operators to define *Hard* resource usage limits that a
*Namespace* may consume.
See [Admission control: Resource Quota](admission_control_resource_quota.md)
### Finalizers
Upon creation of a *Namespace*, the creator may provide a list of *Finalizer* objects.
Upon creation of a *Namespace*, the creator may provide a list of *Finalizer*
objects.
```go
type FinalizerName string
@ -143,13 +156,14 @@ type NamespaceSpec struct {
A *FinalizerName* is a qualified name.
The API Server enforces that a *Namespace* can only be deleted from storage if and only if
it's *Namespace.Spec.Finalizers* is empty.
The API Server enforces that a *Namespace* can only be deleted from storage if
and only if it's *Namespace.Spec.Finalizers* is empty.
A *finalize* operation is the only mechanism to modify the *Namespace.Spec.Finalizers* field post creation.
A *finalize* operation is the only mechanism to modify the
*Namespace.Spec.Finalizers* field post creation.
Each *Namespace* created has *kubernetes* as an item in its list of initial *Namespace.Spec.Finalizers*
set by default.
Each *Namespace* created has *kubernetes* as an item in its list of initial
*Namespace.Spec.Finalizers* set by default.
### Phases
@ -168,39 +182,48 @@ type NamespaceStatus struct {
}
```
A *Namespace* is in the **Active** phase if it does not have a *ObjectMeta.DeletionTimestamp*.
A *Namespace* is in the **Active** phase if it does not have a
*ObjectMeta.DeletionTimestamp*.
A *Namespace* is in the **Terminating** phase if it has a *ObjectMeta.DeletionTimestamp*.
A *Namespace* is in the **Terminating** phase if it has a
*ObjectMeta.DeletionTimestamp*.
**Active**
Upon creation, a *Namespace* goes in the *Active* phase. This means that content may be associated with
a namespace, and all normal interactions with the namespace are allowed to occur in the cluster.
Upon creation, a *Namespace* goes in the *Active* phase. This means that content
may be associated with a namespace, and all normal interactions with the
namespace are allowed to occur in the cluster.
If a DELETE request occurs for a *Namespace*, the *Namespace.ObjectMeta.DeletionTimestamp* is set
to the current server time. A *namespace controller* observes the change, and sets the *Namespace.Status.Phase*
to *Terminating*.
If a DELETE request occurs for a *Namespace*, the
*Namespace.ObjectMeta.DeletionTimestamp* is set to the current server time. A
*namespace controller* observes the change, and sets the
*Namespace.Status.Phase* to *Terminating*.
**Terminating**
A *namespace controller* watches for *Namespace* objects that have a *Namespace.ObjectMeta.DeletionTimestamp*
value set in order to know when to initiate graceful termination of the *Namespace* associated content that
are known to the cluster.
A *namespace controller* watches for *Namespace* objects that have a
*Namespace.ObjectMeta.DeletionTimestamp* value set in order to know when to
initiate graceful termination of the *Namespace* associated content that are
known to the cluster.
The *namespace controller* enumerates each known resource type in that namespace and deletes it one by one.
The *namespace controller* enumerates each known resource type in that namespace
and deletes it one by one.
Admission control blocks creation of new resources in that namespace in order to prevent a race-condition
where the controller could believe all of a given resource type had been deleted from the namespace,
when in fact some other rogue client agent had created new objects. Using admission control in this
scenario allows each of registry implementations for the individual objects to not need to take into account Namespace life-cycle.
Admission control blocks creation of new resources in that namespace in order to
prevent a race-condition where the controller could believe all of a given
resource type had been deleted from the namespace, when in fact some other rogue
client agent had created new objects. Using admission control in this scenario
allows each of registry implementations for the individual objects to not need
to take into account Namespace life-cycle.
Once all objects known to the *namespace controller* have been deleted, the *namespace controller*
executes a *finalize* operation on the namespace that removes the *kubernetes* value from
the *Namespace.Spec.Finalizers* list.
Once all objects known to the *namespace controller* have been deleted, the
*namespace controller* executes a *finalize* operation on the namespace that
removes the *kubernetes* value from the *Namespace.Spec.Finalizers* list.
If the *namespace controller* sees a *Namespace* whose *ObjectMeta.DeletionTimestamp* is set, and
whose *Namespace.Spec.Finalizers* list is empty, it will signal the server to permanently remove
the *Namespace* from storage by sending a final DELETE action to the API server.
If the *namespace controller* sees a *Namespace* whose
*ObjectMeta.DeletionTimestamp* is set, and whose *Namespace.Spec.Finalizers*
list is empty, it will signal the server to permanently remove the *Namespace*
from storage by sending a final DELETE action to the API server.
### REST API
@ -232,15 +255,18 @@ To interact with content associated with a Namespace:
| WATCH | GET | /api/{version}/watch/{resourceType} | Watch for changes to a {resourceType} across all namespaces |
| LIST | GET | /api/{version}/list/{resourceType} | List instances of {resourceType} across all namespaces |
The API server verifies the *Namespace* on resource creation matches the *{namespace}* on the path.
The API server verifies the *Namespace* on resource creation matches the
*{namespace}* on the path.
The API server will associate a resource with a *Namespace* if not populated by the end-user based on the *Namespace* context
of the incoming request. If the *Namespace* of the resource being created, or updated does not match the *Namespace* on the request,
then the API server will reject the request.
The API server will associate a resource with a *Namespace* if not populated by
the end-user based on the *Namespace* context of the incoming request. If the
*Namespace* of the resource being created, or updated does not match the
*Namespace* on the request, then the API server will reject the request.
### Storage
A namespace provides a unique identifier space and therefore must be in the storage path of a resource.
A namespace provides a unique identifier space and therefore must be in the
storage path of a resource.
In etcd, we want to continue to still support efficient WATCH across namespaces.
@ -248,18 +274,19 @@ Resources that persist content in etcd will have storage paths as follows:
/{k8s_storage_prefix}/{resourceType}/{resource.Namespace}/{resource.Name}
This enables consumers to WATCH /registry/{resourceType} for changes across namespace of a particular {resourceType}.
This enables consumers to WATCH /registry/{resourceType} for changes across
namespace of a particular {resourceType}.
### Kubelet
The kubelet will register pod's it sources from a file or http source with a namespace associated with the
*cluster-id*
The kubelet will register pod's it sources from a file or http source with a
namespace associated with the *cluster-id*
### Example: OpenShift Origin managing a Kubernetes Namespace
In this example, we demonstrate how the design allows for agents built on-top of
Kubernetes that manage their own set of resource types associated with a *Namespace*
to take part in Namespace termination.
Kubernetes that manage their own set of resource types associated with a
*Namespace* to take part in Namespace termination.
OpenShift creates a Namespace in Kubernetes
@ -282,9 +309,10 @@ OpenShift creates a Namespace in Kubernetes
}
```
OpenShift then goes and creates a set of resources (pods, services, etc) associated
with the "development" namespace. It also creates its own set of resources in its
own storage associated with the "development" namespace unknown to Kubernetes.
OpenShift then goes and creates a set of resources (pods, services, etc)
associated with the "development" namespace. It also creates its own set of
resources in its own storage associated with the "development" namespace unknown
to Kubernetes.
User deletes the Namespace in Kubernetes, and Namespace now has following state:
@ -308,10 +336,10 @@ User deletes the Namespace in Kubernetes, and Namespace now has following state:
}
```
The Kubernetes *namespace controller* observes the namespace has a *deletionTimestamp*
and begins to terminate all of the content in the namespace that it knows about. Upon
success, it executes a *finalize* action that modifies the *Namespace* by
removing *kubernetes* from the list of finalizers:
The Kubernetes *namespace controller* observes the namespace has a
*deletionTimestamp* and begins to terminate all of the content in the namespace
that it knows about. Upon success, it executes a *finalize* action that modifies
the *Namespace* by removing *kubernetes* from the list of finalizers:
```json
{
@ -333,11 +361,11 @@ removing *kubernetes* from the list of finalizers:
}
```
OpenShift Origin has its own *namespace controller* that is observing cluster state, and
it observes the same namespace had a *deletionTimestamp* assigned to it. It too will go
and purge resources from its own storage that it manages associated with that namespace.
Upon completion, it executes a *finalize* action and removes the reference to "openshift.com/origin"
from the list of finalizers.
OpenShift Origin has its own *namespace controller* that is observing cluster
state, and it observes the same namespace had a *deletionTimestamp* assigned to
it. It too will go and purge resources from its own storage that it manages
associated with that namespace. Upon completion, it executes a *finalize* action
and removes the reference to "openshift.com/origin" from the list of finalizers.
This results in the following state:
@ -361,12 +389,14 @@ This results in the following state:
}
```
At this point, the Kubernetes *namespace controller* in its sync loop will see that the namespace
has a deletion timestamp and that its list of finalizers is empty. As a result, it knows all
content associated from that namespace has been purged. It performs a final DELETE action
to remove that Namespace from the storage.
At this point, the Kubernetes *namespace controller* in its sync loop will see
that the namespace has a deletion timestamp and that its list of finalizers is
empty. As a result, it knows all content associated from that namespace has been
purged. It performs a final DELETE action to remove that Namespace from the
storage.
At this point, all content associated with that Namespace, and the Namespace itself are gone.
At this point, all content associated with that Namespace, and the Namespace
itself are gone.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

40
docs/design/networking.md
View File

@ -44,9 +44,9 @@ There are 4 distinct networking problems to solve:
## Model and motivation
Kubernetes deviates from the default Docker networking model (though as of
Docker 1.8 their network plugins are getting closer). The goal is for each pod
Docker 1.8 their network plugins are getting closer). The goal is for each pod
to have an IP in a flat shared networking namespace that has full communication
with other physical computers and containers across the network. IP-per-pod
with other physical computers and containers across the network. IP-per-pod
creates a clean, backward-compatible model where pods can be treated much like
VMs or physical hosts from the perspectives of port allocation, networking,
naming, service discovery, load balancing, application configuration, and
@ -71,15 +71,15 @@ among other problems.
All containers within a pod behave as if they are on the same host with regard
to networking. They can all reach each other’s ports on localhost. This offers
simplicity (static ports know a priori), security (ports bound to localhost
are visible within the pod but never outside it), and performance. This also
are visible within the pod but never outside it), and performance. This also
reduces friction for applications moving from the world of uncontainerized apps
on physical or virtual hosts. People running application stacks together on
on physical or virtual hosts. People running application stacks together on
the same host have already figured out how to make ports not conflict and have
arranged for clients to find them.
The approach does reduce isolation between containers within a pod &mdash;
ports could conflict, and there can be no container-private ports, but these
seem to be relatively minor issues with plausible future workarounds. Besides,
seem to be relatively minor issues with plausible future workarounds. Besides,
the premise of pods is that containers within a pod share some resources
(volumes, cpu, ram, etc.) and therefore expect and tolerate reduced isolation.
Additionally, the user can control what containers belong to the same pod
@ -88,7 +88,7 @@ whereas, in general, they don't control what pods land together on a host.
## Pod to pod
Because every pod gets a "real" (not machine-private) IP address, pods can
communicate without proxies or translations. The pod can use well-known port
communicate without proxies or translations. The pod can use well-known port
numbers and can avoid the use of higher-level service discovery systems like
DNS-SD, Consul, or Etcd.
@ -98,7 +98,7 @@ each pod has its own IP address that other pods can know. By making IP addresses
and ports the same both inside and outside the pods, we create a NAT-less, flat
address space. Running "ip addr show" should work as expected. This would enable
all existing naming/discovery mechanisms to work out of the box, including
self-registration mechanisms and applications that distribute IP addresses. We
self-registration mechanisms and applications that distribute IP addresses. We
should be optimizing for inter-pod network communication. Within a pod,
containers are more likely to use communication through volumes (e.g., tmpfs) or
IPC.
@ -141,7 +141,7 @@ gcloud compute routes add "${NODE_NAMES[$i]}" \
--next-hop-instance-zone "${ZONE}" &
```
GCE itself does not know anything about these IPs, though. This means that when
GCE itself does not know anything about these IPs, though. This means that when
a pod tries to egress beyond GCE's project the packets must be SNAT'ed
(masqueraded) to the VM's IP, which GCE recognizes and allows.
@ -161,26 +161,26 @@ to serve the purpose outside of GCE.
## Pod to service
The [service](../user-guide/services.md) abstraction provides a way to group pods under a
common access policy (e.g. load-balanced). The implementation of this creates a
common access policy (e.g. load-balanced). The implementation of this creates a
virtual IP which clients can access and which is transparently proxied to the
pods in a Service. Each node runs a kube-proxy process which programs
pods in a Service. Each node runs a kube-proxy process which programs
`iptables` rules to trap access to service IPs and redirect them to the correct
backends. This provides a highly-available load-balancing solution with low
backends. This provides a highly-available load-balancing solution with low
performance overhead by balancing client traffic from a node on that same node.
## External to internal
So far the discussion has been about how to access a pod or service from within
the cluster. Accessing a pod from outside the cluster is a bit more tricky. We
the cluster. Accessing a pod from outside the cluster is a bit more tricky. We
want to offer highly-available, high-performance load balancing to target
Kubernetes Services. Most public cloud providers are simply not flexible enough
Kubernetes Services. Most public cloud providers are simply not flexible enough
yet.
The way this is generally implemented is to set up external load balancers (e.g.
GCE's ForwardingRules or AWS's ELB) which target all nodes in a cluster. When
GCE's ForwardingRules or AWS's ELB) which target all nodes in a cluster. When
traffic arrives at a node it is recognized as being part of a particular Service
and routed to an appropriate backend Pod. This does mean that some traffic will
get double-bounced on the network. Once cloud providers have better offerings
and routed to an appropriate backend Pod. This does mean that some traffic will
get double-bounced on the network. Once cloud providers have better offerings
we can take advantage of those.
## Challenges and future work
@ -207,7 +207,13 @@ External IP assignment would also simplify DNS support (see below).
### IPv6
IPv6 would be a nice option, also, but we can't depend on it yet. Docker support is in progress: [Docker issue #2974](https://github.com/dotcloud/docker/issues/2974), [Docker issue #6923](https://github.com/dotcloud/docker/issues/6923), [Docker issue #6975](https://github.com/dotcloud/docker/issues/6975). Additionally, direct ipv6 assignment to instances doesn't appear to be supported by major cloud providers (e.g., AWS EC2, GCE) yet. We'd happily take pull requests from people running Kubernetes on bare metal, though. :-)
IPv6 would be a nice option, also, but we can't depend on it yet. Docker support
is in progress: [Docker issue #2974](https://github.com/dotcloud/docker/issues/2974),
[Docker issue #6923](https://github.com/dotcloud/docker/issues/6923),
[Docker issue #6975](https://github.com/dotcloud/docker/issues/6975).
Additionally, direct ipv6 assignment to instances doesn't appear to be supported
by major cloud providers (e.g., AWS EC2, GCE) yet. We'd happily take pull
requests from people running Kubernetes on bare metal, though. :-)
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

161
docs/design/nodeaffinity.md
View File

@ -36,40 +36,41 @@ Documentation for other releases can be found at
## Introduction
This document proposes a new label selector representation, called `NodeSelector`,
that is similar in many ways to `LabelSelector`, but is a bit more flexible and is
intended to be used only for selecting nodes.
This document proposes a new label selector representation, called
`NodeSelector`, that is similar in many ways to `LabelSelector`, but is a bit
more flexible and is intended to be used only for selecting nodes.
In addition, we propose to replace the `map[string]string` in `PodSpec` that the scheduler
currently uses as part of restricting the set of nodes onto which a pod is
eligible to schedule, with a field of type `Affinity` that contains one or
more affinity specifications. In this document we discuss `NodeAffinity`, which
contains one or more of the following
In addition, we propose to replace the `map[string]string` in `PodSpec` that the
scheduler currently uses as part of restricting the set of nodes onto which a
pod is eligible to schedule, with a field of type `Affinity` that contains one
or more affinity specifications. In this document we discuss `NodeAffinity`,
which contains one or more of the following:
* a field called `RequiredDuringSchedulingRequiredDuringExecution` that will be
represented by a `NodeSelector`, and thus generalizes the scheduling behavior of
the current `map[string]string` but still serves the purpose of restricting
the set of nodes onto which the pod can schedule. In addition, unlike the behavior
of the current `map[string]string`, when it becomes violated the system will
try to eventually evict the pod from its node.
* a field called `RequiredDuringSchedulingIgnoredDuringExecution` which is identical
to `RequiredDuringSchedulingRequiredDuringExecution` except that the system
may or may not try to eventually evict the pod from its node.
* a field called `PreferredDuringSchedulingIgnoredDuringExecution` that specifies which nodes are
preferred for scheduling among those that meet all scheduling requirements.
the set of nodes onto which the pod can schedule. In addition, unlike the
behavior of the current `map[string]string`, when it becomes violated the system
will try to eventually evict the pod from its node.
* a field called `RequiredDuringSchedulingIgnoredDuringExecution` which is
identical to `RequiredDuringSchedulingRequiredDuringExecution` except that the
system may or may not try to eventually evict the pod from its node.
* a field called `PreferredDuringSchedulingIgnoredDuringExecution` that
specifies which nodes are preferred for scheduling among those that meet all
scheduling requirements.
(In practice, as discussed later, we will actually *add* the `Affinity` field
rather than replacing `map[string]string`, due to backward compatibility requirements.)
rather than replacing `map[string]string`, due to backward compatibility
requirements.)
The affiniy specifications described above allow a pod to request various properties
that are inherent to nodes, for example "run this pod on a node with an Intel CPU" or, in a
multi-zone cluster, "run this pod on a node in zone Z."
The affiniy specifications described above allow a pod to request various
properties that are inherent to nodes, for example "run this pod on a node with
an Intel CPU" or, in a multi-zone cluster, "run this pod on a node in zone Z."
([This issue](https://github.com/kubernetes/kubernetes/issues/9044) describes
some of the properties that a node might publish as labels, which affinity expressions
can match against.)
They do *not* allow a pod to request to schedule
(or not schedule) on a node based on what other pods are running on the node. That
feature is called "inter-pod topological affinity/anti-afinity" and is described
[here](https://github.com/kubernetes/kubernetes/pull/18265).
some of the properties that a node might publish as labels, which affinity
expressions can match against.) They do *not* allow a pod to request to schedule
(or not schedule) on a node based on what other pods are running on the node.
That feature is called "inter-pod topological affinity/anti-afinity" and is
described [here](https://github.com/kubernetes/kubernetes/pull/18265).
## API
@ -171,9 +172,9 @@ type PreferredSchedulingTerm struct {
}
```
Unfortunately, the name of the existing `map[string]string` field in PodSpec is `NodeSelector`
and we can't change it since this name is part of the API. Hopefully this won't
cause too much confusion.
Unfortunately, the name of the existing `map[string]string` field in PodSpec is
`NodeSelector` and we can't change it since this name is part of the API.
Hopefully this won't cause too much confusion.
## Examples
@ -186,81 +187,91 @@ cause too much confusion.
## Backward compatibility
When we add `Affinity` to PodSpec, we will deprecate, but not remove, the current field in PodSpec
When we add `Affinity` to PodSpec, we will deprecate, but not remove, the
current field in PodSpec
```go
NodeSelector map[string]string `json:"nodeSelector,omitempty"`
```
Old version of the scheduler will ignore the `Affinity` field.
New versions of the scheduler will apply their scheduling predicates to both `Affinity` and `nodeSelector`,
i.e. the pod can only schedule onto nodes that satisfy both sets of requirements. We will not
attempt to convert between `Affinity` and `nodeSelector`.
Old version of the scheduler will ignore the `Affinity` field. New versions of
the scheduler will apply their scheduling predicates to both `Affinity` and
`nodeSelector`, i.e. the pod can only schedule onto nodes that satisfy both sets
of requirements. We will not attempt to convert between `Affinity` and
`nodeSelector`.
Old versions of non-scheduling clients will not know how to do anything semantically meaningful
with `Affinity`, but we don't expect that this will cause a problem.
Old versions of non-scheduling clients will not know how to do anything
semantically meaningful with `Affinity`, but we don't expect that this will
cause a problem.
See [this comment](https://github.com/kubernetes/kubernetes/issues/341#issuecomment-140809259)
for more discussion.
Users should not start using `NodeAffinity` until the full implementation has been in Kubelet and the master
for enough binary versions that we feel comfortable that we will not need to roll back either Kubelet
or master to a version that does not support them. Longer-term we will use a programatic approach to
enforcing this (#4855).
Users should not start using `NodeAffinity` until the full implementation has
been in Kubelet and the master for enough binary versions that we feel
comfortable that we will not need to roll back either Kubelet or master to a
version that does not support them. Longer-term we will use a programatic
approach to enforcing this (#4855).
## Implementation plan
1. Add the `Affinity` field to PodSpec and the `NodeAffinity`, `PreferredDuringSchedulingIgnoredDuringExecution`,
and `RequiredDuringSchedulingIgnoredDuringExecution` types to the API
2. Implement a scheduler predicate that takes `RequiredDuringSchedulingIgnoredDuringExecution` into account
3. Implement a scheduler priority function that takes `PreferredDuringSchedulingIgnoredDuringExecution` into account
4. At this point, the feature can be deployed and `PodSpec.NodeSelector` can be marked as deprecated
5. Add the `RequiredDuringSchedulingRequiredDuringExecution` field to the API
6. Modify the scheduler predicate from step 2 to also take `RequiredDuringSchedulingRequiredDuringExecution` into account
7. Add `RequiredDuringSchedulingRequiredDuringExecution` to Kubelet's admission decision
8. Implement code in Kubelet *or* the controllers that evicts a pod that no longer satisfies
`RequiredDuringSchedulingRequiredDuringExecution`
(see [this comment](https://github.com/kubernetes/kubernetes/issues/12744#issuecomment-164372008)).
1. Add the `Affinity` field to PodSpec and the `NodeAffinity`,
`PreferredDuringSchedulingIgnoredDuringExecution`, and
`RequiredDuringSchedulingIgnoredDuringExecution` types to the API.
2. Implement a scheduler predicate that takes
`RequiredDuringSchedulingIgnoredDuringExecution` into account.
3. Implement a scheduler priority function that takes
`PreferredDuringSchedulingIgnoredDuringExecution` into account.
4. At this point, the feature can be deployed and `PodSpec.NodeSelector` can be
marked as deprecated.
5. Add the `RequiredDuringSchedulingRequiredDuringExecution` field to the API.
6. Modify the scheduler predicate from step 2 to also take
`RequiredDuringSchedulingRequiredDuringExecution` into account.
7. Add `RequiredDuringSchedulingRequiredDuringExecution` to Kubelet's admission
decision.
8. Implement code in Kubelet *or* the controllers that evicts a pod that no
longer satisfies `RequiredDuringSchedulingRequiredDuringExecution` (see [this comment](https://github.com/kubernetes/kubernetes/issues/12744#issuecomment-164372008)).
We assume Kubelet publishes labels describing the node's membership in all of the relevant scheduling
domains (e.g. node name, rack name, availability zone name, etc.). See #9044.
We assume Kubelet publishes labels describing the node's membership in all of
the relevant scheduling domains (e.g. node name, rack name, availability zone
name, etc.). See #9044.
## Extensibility
The design described here is the result of careful analysis of use cases, a decade of experience
with Borg at Google, and a review of similar features in other open-source container orchestration
systems. We believe that it properly balances the goal of expressiveness against the goals of
simplicity and efficiency of implementation. However, we recognize that
use cases may arise in the future that cannot be expressed using the syntax described here.
Although we are not implementing an affinity-specific extensibility mechanism for a variety
of reasons (simplicity of the codebase, simplicity of cluster deployment, desire for Kubernetes
users to get a consistent experience, etc.), the regular Kubernetes
annotation mechanism can be used to add or replace affinity rules. The way this work would is
The design described here is the result of careful analysis of use cases, a
decade of experience with Borg at Google, and a review of similar features in
other open-source container orchestration systems. We believe that it properly
balances the goal of expressiveness against the goals of simplicity and
efficiency of implementation. However, we recognize that use cases may arise in
the future that cannot be expressed using the syntax described here. Although we
are not implementing an affinity-specific extensibility mechanism for a variety
of reasons (simplicity of the codebase, simplicity of cluster deployment, desire
for Kubernetes users to get a consistent experience, etc.), the regular
Kubernetes annotation mechanism can be used to add or replace affinity rules.
The way this work would is:
1. Define one or more annotations to describe the new affinity rule(s)
1. User (or an admission controller) attaches the annotation(s) to pods to request the desired scheduling behavior.
If the new rule(s) *replace* one or more fields of `Affinity` then the user would omit those fields
from `Affinity`; if they are *additional rules*, then the user would fill in `Affinity` as well as the
annotation(s).
1. User (or an admission controller) attaches the annotation(s) to pods to
request the desired scheduling behavior. If the new rule(s) *replace* one or
more fields of `Affinity` then the user would omit those fields from `Affinity`;
if they are *additional rules*, then the user would fill in `Affinity` as well
as the annotation(s).
1. Scheduler takes the annotation(s) into account when scheduling.
If some particular new syntax becomes popular, we would consider upstreaming it by integrating
it into the standard `Affinity`.
If some particular new syntax becomes popular, we would consider upstreaming it
by integrating it into the standard `Affinity`.
## Future work
Are there any other fields we should convert from `map[string]string` to `NodeSelector`?
Are there any other fields we should convert from `map[string]string` to
`NodeSelector`?
## Related issues
The review for this proposal is in #18261.
The main related issue is #341. Issue #367 is also related. Those issues reference other
related issues.
The main related issue is #341. Issue #367 is also related. Those issues
reference other related issues.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

96
docs/design/persistent-storage.md
View File

@ -34,43 +34,60 @@ Documentation for other releases can be found at
# Persistent Storage
This document proposes a model for managing persistent, cluster-scoped storage for applications requiring long lived data.
This document proposes a model for managing persistent, cluster-scoped storage
for applications requiring long lived data.
### Abstract
Two new API kinds:
A `PersistentVolume` (PV) is a storage resource provisioned by an administrator. It is analogous to a node. See [Persistent Volume Guide](../user-guide/persistent-volumes/) for how to use it.
A `PersistentVolume` (PV) is a storage resource provisioned by an administrator.
It is analogous to a node. See [Persistent Volume Guide](../user-guide/persistent-volumes/)
for how to use it.
A `PersistentVolumeClaim` (PVC) is a user's request for a persistent volume to use in a pod. It is analogous to a pod.
A `PersistentVolumeClaim` (PVC) is a user's request for a persistent volume to
use in a pod. It is analogous to a pod.
One new system component:
`PersistentVolumeClaimBinder` is a singleton running in master that watches all PersistentVolumeClaims in the system and binds them to the closest matching available PersistentVolume. The volume manager watches the API for newly created volumes to manage.
`PersistentVolumeClaimBinder` is a singleton running in master that watches all
PersistentVolumeClaims in the system and binds them to the closest matching
available PersistentVolume. The volume manager watches the API for newly created
volumes to manage.
One new volume:
`PersistentVolumeClaimVolumeSource` references the user's PVC in the same namespace. This volume finds the bound PV and mounts that volume for the pod. A `PersistentVolumeClaimVolumeSource` is, essentially, a wrapper around another type of volume that is owned by someone else (the system).
`PersistentVolumeClaimVolumeSource` references the user's PVC in the same
namespace. This volume finds the bound PV and mounts that volume for the pod. A
`PersistentVolumeClaimVolumeSource` is, essentially, a wrapper around another
type of volume that is owned by someone else (the system).
Kubernetes makes no guarantees at runtime that the underlying storage exists or is available. High availability is left to the storage provider.
Kubernetes makes no guarantees at runtime that the underlying storage exists or
is available. High availability is left to the storage provider.
### Goals
* Allow administrators to describe available storage
* Allow pod authors to discover and request persistent volumes to use with pods
* Enforce security through access control lists and securing storage to the same namespace as the pod volume
* Enforce quotas through admission control
* Enforce scheduler rules by resource counting
* Ensure developers can rely on storage being available without being closely bound to a particular disk, server, network, or storage device.
* Allow administrators to describe available storage.
* Allow pod authors to discover and request persistent volumes to use with pods.
* Enforce security through access control lists and securing storage to the same
namespace as the pod volume.
* Enforce quotas through admission control.
* Enforce scheduler rules by resource counting.
* Ensure developers can rely on storage being available without being closely
bound to a particular disk, server, network, or storage device.
#### Describe available storage
Cluster administrators use the API to manage *PersistentVolumes*. A custom store `NewPersistentVolumeOrderedIndex` will index volumes by access modes and sort by storage capacity. The `PersistentVolumeClaimBinder` watches for new claims for storage and binds them to an available volume by matching the volume's characteristics (AccessModes and storage size) to the user's request.
Cluster administrators use the API to manage *PersistentVolumes*. A custom store
`NewPersistentVolumeOrderedIndex` will index volumes by access modes and sort by
storage capacity. The `PersistentVolumeClaimBinder` watches for new claims for
storage and binds them to an available volume by matching the volume's
characteristics (AccessModes and storage size) to the user's request.
PVs are system objects and, thus, have no namespace.
Many means of dynamic provisioning will be eventually be implemented for various storage types.
Many means of dynamic provisioning will be eventually be implemented for various
storage types.
##### PersistentVolume API
@ -87,11 +104,15 @@ Many means of dynamic provisioning will be eventually be implemented for various
#### Request Storage
Kubernetes users request persistent storage for their pod by creating a ```PersistentVolumeClaim```. Their request for storage is described by their requirements for resources and mount capabilities.
Kubernetes users request persistent storage for their pod by creating a
```PersistentVolumeClaim```. Their request for storage is described by their
requirements for resources and mount capabilities.
Requests for volumes are bound to available volumes by the volume manager, if a suitable match is found. Requests for resources can go unfulfilled.
Requests for volumes are bound to available volumes by the volume manager, if a
suitable match is found. Requests for resources can go unfulfilled.
Users attach their claim to their pod using a new ```PersistentVolumeClaimVolumeSource``` volume source.
Users attach their claim to their pod using a new
```PersistentVolumeClaimVolumeSource``` volume source.
##### PersistentVolumeClaim API
@ -110,23 +131,31 @@ Users attach their claim to their pod using a new ```PersistentVolumeClaimVolume
#### Scheduling constraints
Scheduling constraints are to be handled similar to pod resource constraints. Pods will need to be annotated or decorated with the number of resources it requires on a node. Similarly, a node will need to list how many it has used or available.
Scheduling constraints are to be handled similar to pod resource constraints.
Pods will need to be annotated or decorated with the number of resources it
requires on a node. Similarly, a node will need to list how many it has used or
available.
TBD
#### Events
The implementation of persistent storage will not require events to communicate to the user the state of their claim. The CLI for bound claims contains a reference to the backing persistent volume. This is always present in the API and CLI, making an event to communicate the same unnecessary.
Events that communicate the state of a mounted volume are left to the volume plugins.
The implementation of persistent storage will not require events to communicate
to the user the state of their claim. The CLI for bound claims contains a
reference to the backing persistent volume. This is always present in the API
and CLI, making an event to communicate the same unnecessary.
Events that communicate the state of a mounted volume are left to the volume
plugins.
### Example
#### Admin provisions storage
An administrator provisions storage by posting PVs to the API. Various way to automate this task can be scripted. Dynamic provisioning is a future feature that can maintain levels of PVs.
An administrator provisions storage by posting PVs to the API. Various ways to
automate this task can be scripted. Dynamic provisioning is a future feature
that can maintain levels of PVs.
```yaml
POST:
@ -152,7 +181,8 @@ pv0001 map[] 10737418240 RWO
#### Users request storage
A user requests storage by posting a PVC to the API. Their request contains the AccessModes they wish their volume to have and the minimum size needed.
A user requests storage by posting a PVC to the API. Their request contains the
AccessModes they wish their volume to have and the minimum size needed.
The user must be within a namespace to create PVCs.
@ -181,7 +211,10 @@ myclaim-1 map[] pending
#### Matching and binding
The ```PersistentVolumeClaimBinder``` attempts to find an available volume that most closely matches the user's request. If one exists, they are bound by putting a reference on the PV to the PVC. Requests can go unfulfilled if a suitable match is not found.
The ```PersistentVolumeClaimBinder``` attempts to find an available volume that
most closely matches the user's request. If one exists, they are bound by
putting a reference on the PV to the PVC. Requests can go unfulfilled if a
suitable match is not found.
```console
$ kubectl get pv
@ -198,9 +231,12 @@ myclaim-1 map[] Bound b16e91d6-c0ef-11e4-8
#### Claim usage
The claim holder can use their claim as a volume. The ```PersistentVolumeClaimVolumeSource``` knows to fetch the PV backing the claim and mount its volume for a pod.
The claim holder can use their claim as a volume. The ```PersistentVolumeClaimVolumeSource``` knows to fetch the PV backing the claim
and mount its volume for a pod.
The claim holder owns the claim and its data for as long as the claim exists. The pod using the claim can be deleted, but the claim remains in the user's namespace. It can be used again and again by many pods.
The claim holder owns the claim and its data for as long as the claim exists.
The pod using the claim can be deleted, but the claim remains in the user's
namespace. It can be used again and again by many pods.
```yaml
POST:
@ -233,9 +269,11 @@ When a claim holder is finished with their data, they can delete their claim.
$ kubectl delete pvc myclaim-1
```
The ```PersistentVolumeClaimBinder``` will reconcile this by removing the claim reference from the PV and change the PVs status to 'Released'.
The ```PersistentVolumeClaimBinder``` will reconcile this by removing the claim
reference from the PV and change the PVs status to 'Released'.
Admins can script the recycling of released volumes. Future dynamic provisioners will understand how a volume should be recycled.
Admins can script the recycling of released volumes. Future dynamic provisioners
will understand how a volume should be recycled.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

699
docs/design/podaffinity.md
View File

@ -38,45 +38,48 @@ Documentation for other releases can be found at
NOTE: It is useful to read about [node affinity](nodeaffinity.md) first.
This document describes a proposal for specifying and implementing inter-pod topological affinity and
anti-affinity. By that we mean: rules that specify that certain pods should be placed
in the same topological domain (e.g. same node, same rack, same zone, same
power domain, etc.) as some other pods, or, conversely, should *not* be placed in the
same topological domain as some other pods.
This document describes a proposal for specifying and implementing inter-pod
topological affinity and anti-affinity. By that we mean: rules that specify that
certain pods should be placed in the same topological domain (e.g. same node,
same rack, same zone, same power domain, etc.) as some other pods, or,
conversely, should *not* be placed in the same topological domain as some other
pods.
Here are a few example rules; we explain how to express them using the API described
in this doc later, in the section "Examples."
Here are a few example rules; we explain how to express them using the API
described in this doc later, in the section "Examples."
* Affinity
* Co-locate the pods from a particular service or Job in the same availability zone,
without specifying which zone that should be.
* Co-locate the pods from service S1 with pods from service S2 because S1 uses S2
and thus it is useful to minimize the network latency between them. Co-location
might mean same nodes and/or same availability zone.
* Co-locate the pods from a particular service or Job in the same availability
zone, without specifying which zone that should be.
* Co-locate the pods from service S1 with pods from service S2 because S1 uses
S2 and thus it is useful to minimize the network latency between them.
Co-location might mean same nodes and/or same availability zone.
* Anti-affinity
* Spread the pods of a service across nodes and/or availability zones,
e.g. to reduce correlated failures
* Give a pod "exclusive" access to a node to guarantee resource isolation -- it must never share the node with other pods
* Spread the pods of a service across nodes and/or availability zones, e.g. to
reduce correlated failures.
* Give a pod "exclusive" access to a node to guarantee resource isolation --
it must never share the node with other pods.
* Don't schedule the pods of a particular service on the same nodes as pods of
another service that are known to interfere with the performance of the pods of the first service.
another service that are known to interfere with the performance of the pods of
the first service.
For both affinity and anti-affinity, there are three variants. Two variants have the
property of requiring the affinity/anti-affinity to be satisfied for the pod to be allowed
to schedule onto a node; the difference between them is that if the condition ceases to
be met later on at runtime, for one of them the system will try to eventually evict the pod,
while for the other the system may not try to do so. The third variant
simply provides scheduling-time *hints* that the scheduler will try
to satisfy but may not be able to. These three variants are directly analogous to the three
variants of [node affinity](nodeaffinity.md).
For both affinity and anti-affinity, there are three variants. Two variants have
the property of requiring the affinity/anti-affinity to be satisfied for the pod
to be allowed to schedule onto a node; the difference between them is that if
the condition ceases to be met later on at runtime, for one of them the system
will try to eventually evict the pod, while for the other the system may not try
to do so. The third variant simply provides scheduling-time *hints* that the
scheduler will try to satisfy but may not be able to. These three variants are
directly analogous to the three variants of [node affinity](nodeaffinity.md).
Note that this proposal is only about *inter-pod* topological affinity and anti-affinity.
There are other forms of topological affinity and anti-affinity. For example,
you can use [node affinity](nodeaffinity.md) to require (prefer)
that a set of pods all be scheduled in some specific zone Z. Node affinity is not
capable of expressing inter-pod dependencies, and conversely the API
we describe in this document is not capable of expressing node affinity rules.
For simplicity, we will use the terms "affinity" and "anti-affinity" to mean
"inter-pod topological affinity" and "inter-pod topological anti-affinity," respectively,
in the remainder of this document.
Note that this proposal is only about *inter-pod* topological affinity and
anti-affinity. There are other forms of topological affinity and anti-affinity.
For example, you can use [node affinity](nodeaffinity.md) to require (prefer)
that a set of pods all be scheduled in some specific zone Z. Node affinity is
not capable of expressing inter-pod dependencies, and conversely the API we
describe in this document is not capable of expressing node affinity rules. For
simplicity, we will use the terms "affinity" and "anti-affinity" to mean
"inter-pod topological affinity" and "inter-pod topological anti-affinity,"
respectively, in the remainder of this document.
## API
@ -90,28 +93,28 @@ The `Affinity` type is defined as follows
```go
type Affinity struct {
PodAffinity *PodAffinity `json:"podAffinity,omitempty"`
PodAntiAffinity *PodAntiAffinity `json:"podAntiAffinity,omitempty"`
PodAffinity *PodAffinity `json:"podAffinity,omitempty"`
PodAntiAffinity *PodAntiAffinity `json:"podAntiAffinity,omitempty"`
}
type PodAffinity struct {
// If the affinity requirements specified by this field are not met at
// If the affinity requirements specified by this field are not met at
// scheduling time, the pod will not be scheduled onto the node.
// If the affinity requirements specified by this field cease to be met
// at some point during pod execution (e.g. due to a pod label update), the
// system will try to eventually evict the pod from its node.
// When there are multiple elements, the lists of nodes corresponding to each
// PodAffinityTerm are intersected, i.e. all terms must be satisfied.
RequiredDuringSchedulingRequiredDuringExecution []PodAffinityTerm `json:"requiredDuringSchedulingRequiredDuringExecution,omitempty"`
// When there are multiple elements, the lists of nodes corresponding to each
// PodAffinityTerm are intersected, i.e. all terms must be satisfied.
RequiredDuringSchedulingRequiredDuringExecution []PodAffinityTerm `json:"requiredDuringSchedulingRequiredDuringExecution,omitempty"`
// If the affinity requirements specified by this field are not met at
// scheduling time, the pod will not be scheduled onto the node.
// If the affinity requirements specified by this field cease to be met
// at some point during pod execution (e.g. due to a pod label update), the
// system may or may not try to eventually evict the pod from its node.
// When there are multiple elements, the lists of nodes corresponding to each
// PodAffinityTerm are intersected, i.e. all terms must be satisfied.
RequiredDuringSchedulingIgnoredDuringExecution []PodAffinityTerm `json:"requiredDuringSchedulingIgnoredDuringExecution,omitempty"`
// The scheduler will prefer to schedule pods to nodes that satisfy
// When there are multiple elements, the lists of nodes corresponding to each
// PodAffinityTerm are intersected, i.e. all terms must be satisfied.
RequiredDuringSchedulingIgnoredDuringExecution []PodAffinityTerm `json:"requiredDuringSchedulingIgnoredDuringExecution,omitempty"`
// The scheduler will prefer to schedule pods to nodes that satisfy
// the affinity expressions specified by this field, but it may choose
// a node that violates one or more of the expressions. The node that is
// most preferred is the one with the greatest sum of weights, i.e.
@ -120,27 +123,27 @@ type PodAffinity struct {
// compute a sum by iterating through the elements of this field and adding
// "weight" to the sum if the node matches the corresponding MatchExpressions; the
// node(s) with the highest sum are the most preferred.
PreferredDuringSchedulingIgnoredDuringExecution []WeightedPodAffinityTerm `json:"preferredDuringSchedulingIgnoredDuringExecution,omitempty"`
PreferredDuringSchedulingIgnoredDuringExecution []WeightedPodAffinityTerm `json:"preferredDuringSchedulingIgnoredDuringExecution,omitempty"`
}
type PodAntiAffinity struct {
// If the anti-affinity requirements specified by this field are not met at
// If the anti-affinity requirements specified by this field are not met at
// scheduling time, the pod will not be scheduled onto the node.
// If the anti-affinity requirements specified by this field cease to be met
// at some point during pod execution (e.g. due to a pod label update), the
// system will try to eventually evict the pod from its node.
// When there are multiple elements, the lists of nodes corresponding to each
// PodAffinityTerm are intersected, i.e. all terms must be satisfied.
RequiredDuringSchedulingRequiredDuringExecution []PodAffinityTerm `json:"requiredDuringSchedulingRequiredDuringExecution,omitempty"`
// When there are multiple elements, the lists of nodes corresponding to each
// PodAffinityTerm are intersected, i.e. all terms must be satisfied.
RequiredDuringSchedulingRequiredDuringExecution []PodAffinityTerm `json:"requiredDuringSchedulingRequiredDuringExecution,omitempty"`
// If the anti-affinity requirements specified by this field are not met at
// scheduling time, the pod will not be scheduled onto the node.
// If the anti-affinity requirements specified by this field cease to be met
// at some point during pod execution (e.g. due to a pod label update), the
// system may or may not try to eventually evict the pod from its node.
// When there are multiple elements, the lists of nodes corresponding to each
// PodAffinityTerm are intersected, i.e. all terms must be satisfied.
RequiredDuringSchedulingIgnoredDuringExecution []PodAffinityTerm `json:"requiredDuringSchedulingIgnoredDuringExecution,omitempty"`
// The scheduler will prefer to schedule pods to nodes that satisfy
// When there are multiple elements, the lists of nodes corresponding to each
// PodAffinityTerm are intersected, i.e. all terms must be satisfied.
RequiredDuringSchedulingIgnoredDuringExecution []PodAffinityTerm `json:"requiredDuringSchedulingIgnoredDuringExecution,omitempty"`
// The scheduler will prefer to schedule pods to nodes that satisfy
// the anti-affinity expressions specified by this field, but it may choose
// a node that violates one or more of the expressions. The node that is
// most preferred is the one with the greatest sum of weights, i.e.
@ -149,7 +152,7 @@ type PodAntiAffinity struct {
// compute a sum by iterating through the elements of this field and adding
// "weight" to the sum if the node matches the corresponding MatchExpressions; the
// node(s) with the highest sum are the most preferred.
PreferredDuringSchedulingIgnoredDuringExecution []WeightedPodAffinityTerm `json:"preferredDuringSchedulingIgnoredDuringExecution,omitempty"`
PreferredDuringSchedulingIgnoredDuringExecution []WeightedPodAffinityTerm `json:"preferredDuringSchedulingIgnoredDuringExecution,omitempty"`
}
type WeightedPodAffinityTerm struct {
@ -159,23 +162,25 @@ type WeightedPodAffinityTerm struct {
}
type PodAffinityTerm struct {
LabelSelector *LabelSelector `json:"labelSelector,omitempty"`
// namespaces specifies which namespaces the LabelSelector applies to (matches against);
// nil list means "this pod's namespace," empty list means "all namespaces"
// The json tag here is not "omitempty" since we need to distinguish nil and empty.
// See https://golang.org/pkg/encoding/json/#Marshal for more details.
Namespaces []api.Namespace `json:"namespaces,omitempty"`
// empty topology key is interpreted by the scheduler as "all topologies"
TopologyKey string `json:"topologyKey,omitempty"`
LabelSelector *LabelSelector `json:"labelSelector,omitempty"`
// namespaces specifies which namespaces the LabelSelector applies to (matches against);
// nil list means "this pod's namespace," empty list means "all namespaces"
// The json tag here is not "omitempty" since we need to distinguish nil and empty.
// See https://golang.org/pkg/encoding/json/#Marshal for more details.
Namespaces []api.Namespace `json:"namespaces,omitempty"`
// empty topology key is interpreted by the scheduler as "all topologies"
TopologyKey string `json:"topologyKey,omitempty"`
}
```
Note that the `Namespaces` field is necessary because normal `LabelSelector` is scoped
to the pod's namespace, but we need to be able to match against all pods globally.
Note that the `Namespaces` field is necessary because normal `LabelSelector` is
scoped to the pod's namespace, but we need to be able to match against all pods
globally.
To explain how this API works, let's say that the `PodSpec` of a pod `P` has an `Affinity`
that is configured as follows (note that we've omitted and collapsed some fields for
simplicity, but this should sufficiently convey the intent of the design):
To explain how this API works, let's say that the `PodSpec` of a pod `P` has an
`Affinity` that is configured as follows (note that we've omitted and collapsed
some fields for simplicity, but this should sufficiently convey the intent of
the design):
```go
PodAffinity {
@ -188,130 +193,160 @@ PodAntiAffinity {
}
```
Then when scheduling pod P, the scheduler
* Can only schedule P onto nodes that are running pods that satisfy `P1`. (Assumes all nodes have a label with key `node` and value specifying their node name.)
* Should try to schedule P onto zones that are running pods that satisfy `P2`. (Assumes all nodes have a label with key `zone` and value specifying their zone.)
* Cannot schedule P onto any racks that are running pods that satisfy `P3`. (Assumes all nodes have a label with key `rack` and value specifying their rack name.)
* Should try not to schedule P onto any power domains that are running pods that satisfy `P4`. (Assumes all nodes have a label with key `power` and value specifying their power domain.)
Then when scheduling pod P, the scheduler:
* Can only schedule P onto nodes that are running pods that satisfy `P1`.
(Assumes all nodes have a label with key `node` and value specifying their node
name.)
* Should try to schedule P onto zones that are running pods that satisfy `P2`.
(Assumes all nodes have a label with key `zone` and value specifying their
zone.)
* Cannot schedule P onto any racks that are running pods that satisfy `P3`.
(Assumes all nodes have a label with key `rack` and value specifying their rack
name.)
* Should try not to schedule P onto any power domains that are running pods that
satisfy `P4`. (Assumes all nodes have a label with key `power` and value
specifying their power domain.)
When `RequiredDuringScheduling` has multiple elements, the requirements are ANDed.
For `PreferredDuringScheduling` the weights are added for the terms that are satisfied for each node, and
the node(s) with the highest weight(s) are the most preferred.
When `RequiredDuringScheduling` has multiple elements, the requirements are
ANDed. For `PreferredDuringScheduling` the weights are added for the terms that
are satisfied for each node, and the node(s) with the highest weight(s) are the
most preferred.
In reality there are two variants of `RequiredDuringScheduling`: one suffixed with
`RequiredDuringEecution` and one suffixed with `IgnoredDuringExecution`. For the
first variant, if the affinity/anti-affinity ceases to be met at some point during
pod execution (e.g. due to a pod label update), the system will try to eventually evict the pod
from its node. In the second variant, the system may or may not try to eventually
evict the pod from its node.
In reality there are two variants of `RequiredDuringScheduling`: one suffixed
with `RequiredDuringEecution` and one suffixed with `IgnoredDuringExecution`.
For the first variant, if the affinity/anti-affinity ceases to be met at some
point during pod execution (e.g. due to a pod label update), the system will try
to eventually evict the pod from its node. In the second variant, the system may
or may not try to eventually evict the pod from its node.
## A comment on symmetry
One thing that makes affinity and anti-affinity tricky is symmetry.
Imagine a cluster that is running pods from two services, S1 and S2. Imagine that the pods of S1 have a RequiredDuringScheduling anti-affinity rule
"do not run me on nodes that are running pods from S2." It is not sufficient just to check that there are no S2 pods on a node when
you are scheduling a S1 pod. You also need to ensure that there are no S1 pods on a node when you are scheduling a S2 pod,
*even though the S2 pod does not have any anti-affinity rules*. Otherwise if an S1 pod schedules before an S2 pod, the S1
pod's RequiredDuringScheduling anti-affinity rule can be violated by a later-arriving S2 pod. More specifically, if S1 has the aforementioned
RequiredDuringScheduling anti-affinity rule, then
Imagine a cluster that is running pods from two services, S1 and S2. Imagine
that the pods of S1 have a RequiredDuringScheduling anti-affinity rule "do not
run me on nodes that are running pods from S2." It is not sufficient just to
check that there are no S2 pods on a node when you are scheduling a S1 pod. You
also need to ensure that there are no S1 pods on a node when you are scheduling
a S2 pod, *even though the S2 pod does not have any anti-affinity rules*.
Otherwise if an S1 pod schedules before an S2 pod, the S1 pod's
RequiredDuringScheduling anti-affinity rule can be violated by a later-arriving
S2 pod. More specifically, if S1 has the aforementioned RequiredDuringScheduling
anti-affinity rule, then:
* if a node is empty, you can schedule S1 or S2 onto the node
* if a node is running S1 (S2), you cannot schedule S2 (S1) onto the node
Note that while RequiredDuringScheduling anti-affinity is symmetric,
RequiredDuringScheduling affinity is *not* symmetric. That is, if the pods of S1 have a RequiredDuringScheduling affinity rule "run me on nodes that are running
pods from S2," it is not required that there be S1 pods on a node in order to schedule a S2 pod onto that node. More
specifically, if S1 has the aforementioned RequiredDuringScheduling affinity rule, then
RequiredDuringScheduling affinity is *not* symmetric. That is, if the pods of S1
have a RequiredDuringScheduling affinity rule "run me on nodes that are running
pods from S2," it is not required that there be S1 pods on a node in order to
schedule a S2 pod onto that node. More specifically, if S1 has the
aforementioned RequiredDuringScheduling affinity rule, then:
* if a node is empty, you can schedule S2 onto the node
* if a node is empty, you cannot schedule S1 onto the node
* if a node is running S2, you can schedule S1 onto the node
* if a node is running S1+S2 and S1 terminates, S2 continues running
* if a node is running S1+S2 and S2 terminates, the system terminates S1 (eventually)
* if a node is running S1+S2 and S2 terminates, the system terminates S1
(eventually)
However, although RequiredDuringScheduling affinity is not symmetric, there is an implicit PreferredDuringScheduling affinity rule corresponding to every
RequiredDuringScheduling affinity rule: if the pods of S1 have a RequiredDuringScheduling affinity rule "run me on nodes that are running
pods from S2" then it is not required that there be S1 pods on a node in order to schedule a S2 pod onto that node,
but it would be better if there are.
However, although RequiredDuringScheduling affinity is not symmetric, there is
an implicit PreferredDuringScheduling affinity rule corresponding to every
RequiredDuringScheduling affinity rule: if the pods of S1 have a
RequiredDuringScheduling affinity rule "run me on nodes that are running pods
from S2" then it is not required that there be S1 pods on a node in order to
schedule a S2 pod onto that node, but it would be better if there are.
PreferredDuringScheduling is symmetric.
If the pods of S1 had a PreferredDuringScheduling anti-affinity rule "try not to run me on nodes that are running pods from S2"
then we would prefer to keep a S1 pod that we are scheduling off of nodes that are running S2 pods, and also
to keep a S2 pod that we are scheduling off of nodes that are running S1 pods. Likewise if the pods of
S1 had a PreferredDuringScheduling affinity rule "try to run me on nodes that are running pods from S2" then we would prefer
to place a S1 pod that we are scheduling onto a node that is running a S2 pod, and also to place
a S2 pod that we are scheduling onto a node that is running a S1 pod.
PreferredDuringScheduling is symmetric. If the pods of S1 had a
PreferredDuringScheduling anti-affinity rule "try not to run me on nodes that
are running pods from S2" then we would prefer to keep a S1 pod that we are
scheduling off of nodes that are running S2 pods, and also to keep a S2 pod that
we are scheduling off of nodes that are running S1 pods. Likewise if the pods of
S1 had a PreferredDuringScheduling affinity rule "try to run me on nodes that
are running pods from S2" then we would prefer to place a S1 pod that we are
scheduling onto a node that is running a S2 pod, and also to place a S2 pod that
we are scheduling onto a node that is running a S1 pod.
## Examples
Here are some examples of how you would express various affinity and anti-affinity rules using the API we described.
Here are some examples of how you would express various affinity and
anti-affinity rules using the API we described.
### Affinity
In the examples below, the word "put" is intentionally ambiguous; the rules are the same
whether "put" means "must put" (RequiredDuringScheduling) or "try to put"
(PreferredDuringScheduling)--all that changes is which field the rule goes into.
Also, we only discuss scheduling-time, and ignore the execution-time.
Finally, some of the examples
use "zone" and some use "node," just to make the examples more interesting; any of the examples
with "zone" will also work for "node" if you change the `TopologyKey`, and vice-versa.
In the examples below, the word "put" is intentionally ambiguous; the rules are
the same whether "put" means "must put" (RequiredDuringScheduling) or "try to
put" (PreferredDuringScheduling)--all that changes is which field the rule goes
into. Also, we only discuss scheduling-time, and ignore the execution-time.
Finally, some of the examples use "zone" and some use "node," just to make the
examples more interesting; any of the examples with "zone" will also work for
"node" if you change the `TopologyKey`, and vice-versa.
* **Put the pod in zone Z**:
Tricked you! It is not possible express this using the API described here. For this you should use node affinity.
Tricked you! It is not possible express this using the API described here. For
this you should use node affinity.
* **Put the pod in a zone that is running at least one pod from service S**:
`{LabelSelector: <selector that matches S's pods>, TopologyKey: "zone"}`
* **Put the pod on a node that is already running a pod that requires a license for software package P**:
Assuming pods that require a license for software package P have a label `{key=license, value=P}`:
* **Put the pod on a node that is already running a pod that requires a license
for software package P**: Assuming pods that require a license for software
package P have a label `{key=license, value=P}`:
`{LabelSelector: "license" In "P", TopologyKey: "node"}`
* **Put this pod in the same zone as other pods from its same service**:
Assuming pods from this pod's service have some label `{key=service, value=S}`:
`{LabelSelector: "service" In "S", TopologyKey: "zone"}`
This last example illustrates a small issue with this API when it is used
with a scheduler that processes the pending queue one pod at a time, like the current
This last example illustrates a small issue with this API when it is used with a
scheduler that processes the pending queue one pod at a time, like the current
Kubernetes scheduler. The RequiredDuringScheduling rule
`{LabelSelector: "service" In "S", TopologyKey: "zone"}`
only "works" once one pod from service S has been scheduled. But if all pods in service
S have this RequiredDuringScheduling rule in their PodSpec, then the RequiredDuringScheduling rule
will block the first
pod of the service from ever scheduling, since it is only allowed to run in a zone with another pod from
the same service. And of course that means none of the pods of the service will be able
to schedule. This problem *only* applies to RequiredDuringScheduling affinity, not
PreferredDuringScheduling affinity or any variant of anti-affinity.
There are at least three ways to solve this problem
* **short-term**: have the scheduler use a rule that if the RequiredDuringScheduling affinity requirement
matches a pod's own labels, and there are no other such pods anywhere, then disregard the requirement.
This approach has a corner case when running parallel schedulers that are allowed to
schedule pods from the same replicated set (e.g. a single PodTemplate): both schedulers may try to
schedule pods from the set
at the same time and think there are no other pods from that set scheduled yet (e.g. they are
trying to schedule the first two pods from the set), but by the time
the second binding is committed, the first one has already been committed, leaving you with
two pods running that do not respect their RequiredDuringScheduling affinity. There is no
simple way to detect this "conflict" at scheduling time given the current system implementation.
* **longer-term**: when a controller creates pods from a PodTemplate, for exactly *one* of those
pods, it should omit any RequiredDuringScheduling affinity rules that select the pods of that PodTemplate.
* **very long-term/speculative**: controllers could present the scheduler with a group of pods from
the same PodTemplate as a single unit. This is similar to the first approach described above but
avoids the corner case. No special logic is needed in the controllers. Moreover, this would allow
the scheduler to do proper [gang scheduling](https://github.com/kubernetes/kubernetes/issues/16845)
since it could receive an entire gang simultaneously as a single unit.
only "works" once one pod from service S has been scheduled. But if all pods in
service S have this RequiredDuringScheduling rule in their PodSpec, then the
RequiredDuringScheduling rule will block the first pod of the service from ever
scheduling, since it is only allowed to run in a zone with another pod from the
same service. And of course that means none of the pods of the service will be
able to schedule. This problem *only* applies to RequiredDuringScheduling
affinity, not PreferredDuringScheduling affinity or any variant of
anti-affinity. There are at least three ways to solve this problem:
* **short-term**: have the scheduler use a rule that if the
RequiredDuringScheduling affinity requirement matches a pod's own labels, and
there are no other such pods anywhere, then disregard the requirement. This
approach has a corner case when running parallel schedulers that are allowed to
schedule pods from the same replicated set (e.g. a single PodTemplate): both
schedulers may try to schedule pods from the set at the same time and think
there are no other pods from that set scheduled yet (e.g. they are trying to
schedule the first two pods from the set), but by the time the second binding is
committed, the first one has already been committed, leaving you with two pods
running that do not respect their RequiredDuringScheduling affinity. There is no
simple way to detect this "conflict" at scheduling time given the current system
implementation.
* **longer-term**: when a controller creates pods from a PodTemplate, for
exactly *one* of those pods, it should omit any RequiredDuringScheduling
affinity rules that select the pods of that PodTemplate.
* **very long-term/speculative**: controllers could present the scheduler with a
group of pods from the same PodTemplate as a single unit. This is similar to the
first approach described above but avoids the corner case. No special logic is
needed in the controllers. Moreover, this would allow the scheduler to do proper
[gang scheduling](https://github.com/kubernetes/kubernetes/issues/16845) since
it could receive an entire gang simultaneously as a single unit.
### Anti-affinity
As with the affinity examples, the examples here can be RequiredDuringScheduling or
PreferredDuringScheduling anti-affinity, i.e.
"don't" can be interpreted as "must not" or as "try not to" depending on whether the rule appears
in `RequiredDuringScheduling` or `PreferredDuringScheduling`.
As with the affinity examples, the examples here can be RequiredDuringScheduling
or PreferredDuringScheduling anti-affinity, i.e. "don't" can be interpreted as
"must not" or as "try not to" depending on whether the rule appears in
`RequiredDuringScheduling` or `PreferredDuringScheduling`.
* **Spread the pods of this service S across nodes and zones**:
`{{LabelSelector: <selector that matches S's pods>, TopologyKey: "node"}, {LabelSelector: <selector that matches S's pods>, TopologyKey: "zone"}}`
(note that if this is specified as a RequiredDuringScheduling anti-affinity, then the first clause is redundant, since the second
clause will force the scheduler to not put more than one pod from S in the same zone, and thus by
definition it will not put more than one pod from S on the same node, assuming each node is in one zone.
This rule is more useful as PreferredDuringScheduling anti-affinity, e.g. one might expect it to be common in
`{{LabelSelector: <selector that matches S's pods>, TopologyKey: "node"},
{LabelSelector: <selector that matches S's pods>, TopologyKey: "zone"}}`
(note that if this is specified as a RequiredDuringScheduling anti-affinity,
then the first clause is redundant, since the second clause will force the
scheduler to not put more than one pod from S in the same zone, and thus by
definition it will not put more than one pod from S on the same node, assuming
each node is in one zone. This rule is more useful as PreferredDuringScheduling
anti-affinity, e.g. one might expect it to be common in
[Ubernetes](../../docs/proposals/federation.md) clusters.)
* **Don't co-locate pods of this service with pods from service "evilService"**:
@ -323,25 +358,29 @@ This rule is more useful as PreferredDuringScheduling anti-affinity, e.g. one mi
* **Don't co-locate pods of this service with any other pods except other pods of this service**:
Assuming pods from the service have some label `{key=service, value=S}`:
`{LabelSelector: "service" NotIn "S", TopologyKey: "node"}`
Note that this works because `"service" NotIn "S"` matches pods with no key "service"
as well as pods with key "service" and a corresponding value that is not "S."
Note that this works because `"service" NotIn "S"` matches pods with no key
"service" as well as pods with key "service" and a corresponding value that is
not "S."
## Algorithm
An example algorithm a scheduler might use to implement affinity and anti-affinity rules is as follows.
There are certainly more efficient ways to do it; this is just intended to demonstrate that the API's
semantics are implementable.
An example algorithm a scheduler might use to implement affinity and
anti-affinity rules is as follows. There are certainly more efficient ways to
do it; this is just intended to demonstrate that the API's semantics are
implementable.
Terminology definition: We say a pod P is "feasible" on a node N if P meets all of the scheduler
predicates for scheduling P onto N. Note that this algorithm is only concerned about scheduling
time, thus it makes no distinction between RequiredDuringExecution and IgnoredDuringExecution.
Terminology definition: We say a pod P is "feasible" on a node N if P meets all
of the scheduler predicates for scheduling P onto N. Note that this algorithm is
only concerned about scheduling time, thus it makes no distinction between
RequiredDuringExecution and IgnoredDuringExecution.
To make the algorithm slightly more readable, we use the term "HardPodAffinity" as shorthand
for "RequiredDuringSchedulingScheduling pod affinity" and "SoftPodAffinity" as shorthand for
"PreferredDuringScheduling pod affinity." Analogously for "HardPodAntiAffinity" and "SoftPodAntiAffinity."
To make the algorithm slightly more readable, we use the term "HardPodAffinity"
as shorthand for "RequiredDuringSchedulingScheduling pod affinity" and
"SoftPodAffinity" as shorthand for "PreferredDuringScheduling pod affinity."
Analogously for "HardPodAntiAffinity" and "SoftPodAntiAffinity."
** TODO: Update this algorithm to take weight for SoftPod{Affinity,AntiAffinity} into account;
currently it assumes all terms have weight 1. **
** TODO: Update this algorithm to take weight for SoftPod{Affinity,AntiAffinity}
into account; currently it assumes all terms have weight 1. **
```
Z = the pod you are scheduling
@ -389,74 +428,81 @@ foreach node A of {N}
## Special considerations for RequiredDuringScheduling anti-affinity
In this section we discuss three issues with RequiredDuringScheduling anti-affinity:
Denial of Service (DoS), co-existing with daemons, and determining which pod(s) to kill.
See issue #18265 for additional discussion of these topics.
In this section we discuss three issues with RequiredDuringScheduling
anti-affinity: Denial of Service (DoS), co-existing with daemons, and
determining which pod(s) to kill. See issue #18265 for additional discussion of
these topics.
### Denial of Service
Without proper safeguards, a pod using RequiredDuringScheduling anti-affinity can intentionally
or unintentionally cause various problems for other pods, due to the symmetry property of anti-affinity.
Without proper safeguards, a pod using RequiredDuringScheduling anti-affinity
can intentionally or unintentionally cause various problems for other pods, due
to the symmetry property of anti-affinity.
The most notable danger is the ability for a
pod that arrives first to some topology domain, to block all other pods from
scheduling there by stating a conflict with all other pods.
The standard approach
to preventing resource hogging is quota, but simple resource quota cannot prevent
this scenario because the pod may request very little resources. Addressing this
using quota requires a quota scheme that charges based on "opportunity cost" rather
than based simply on requested resources. For example, when handling a pod that expresses
The most notable danger is the ability for a pod that arrives first to some
topology domain, to block all other pods from scheduling there by stating a
conflict with all other pods. The standard approach to preventing resource
hogging is quota, but simple resource quota cannot prevent this scenario because
the pod may request very little resources. Addressing this using quota requires
a quota scheme that charges based on "opportunity cost" rather than based simply
on requested resources. For example, when handling a pod that expresses
RequiredDuringScheduling anti-affinity for all pods using a "node" `TopologyKey`
(i.e. exclusive access to a node), it could charge for the resources of the
average or largest node in the cluster. Likewise if a pod expresses RequiredDuringScheduling
anti-affinity for all pods using a "cluster" `TopologyKey`, it could charge for the resources of the
entire cluster. If node affinity is used to
constrain the pod to a particular topology domain, then the admission-time quota
charging should take that into account (e.g. not charge for the average/largest machine
if the PodSpec constrains the pod to a specific machine with a known size; instead charge
for the size of the actual machine that the pod was constrained to). In all cases
once the pod is scheduled, the quota charge should be adjusted down to the
actual amount of resources allocated (e.g. the size of the actual machine that was
assigned, not the average/largest). If a cluster administrator wants to overcommit quota, for
average or largest node in the cluster. Likewise if a pod expresses
RequiredDuringScheduling anti-affinity for all pods using a "cluster"
`TopologyKey`, it could charge for the resources of the entire cluster. If node
affinity is used to constrain the pod to a particular topology domain, then the
admission-time quota charging should take that into account (e.g. not charge for
the average/largest machine if the PodSpec constrains the pod to a specific
machine with a known size; instead charge for the size of the actual machine
that the pod was constrained to). In all cases once the pod is scheduled, the
quota charge should be adjusted down to the actual amount of resources allocated
(e.g. the size of the actual machine that was assigned, not the
average/largest). If a cluster administrator wants to overcommit quota, for
example to allow more than N pods across all users to request exclusive node
access in a cluster with N nodes, then a priority/preemption scheme should be added
so that the most important pods run when resource demand exceeds supply.
access in a cluster with N nodes, then a priority/preemption scheme should be
added so that the most important pods run when resource demand exceeds supply.
An alternative approach, which is a bit of a blunt hammer, is to use a
capability mechanism to restrict use of RequiredDuringScheduling anti-affinity
to trusted users. A more complex capability mechanism might only restrict it when
using a non-"node" TopologyKey.
to trusted users. A more complex capability mechanism might only restrict it
when using a non-"node" TopologyKey.
Our initial implementation will use a variant of the capability approach, which
requires no configuration: we will simply reject ALL requests, regardless of user,
that specify "all namespaces" with non-"node" TopologyKey for RequiredDuringScheduling anti-affinity.
This allows the "exclusive node" use case while prohibiting the more dangerous ones.
requires no configuration: we will simply reject ALL requests, regardless of
user, that specify "all namespaces" with non-"node" TopologyKey for
RequiredDuringScheduling anti-affinity. This allows the "exclusive node" use
case while prohibiting the more dangerous ones.
A weaker variant of the problem described in the previous paragraph is a pod's ability to use anti-affinity to degrade
the scheduling quality of another pod, but not completely block it from scheduling.
For example, a set of pods S1 could use node affinity to request to schedule onto a set
of nodes that some other set of pods S2 prefers to schedule onto. If the pods in S1
have RequiredDuringScheduling or even PreferredDuringScheduling pod anti-affinity for S2,
then due to the symmetry property of anti-affinity, they can prevent the pods in S2 from
scheduling onto their preferred nodes if they arrive first (for sure in the RequiredDuringScheduling case, and
with some probability that depends on the weighting scheme for the PreferredDuringScheduling case).
A very sophisticated priority and/or quota scheme could mitigate this, or alternatively
we could eliminate the symmetry property of the implementation of PreferredDuringScheduling anti-affinity.
Then only RequiredDuringScheduling anti-affinity could affect scheduling quality
of another pod, and as we described in the previous paragraph, such pods could be charged
quota for the full topology domain, thereby reducing the potential for abuse.
A weaker variant of the problem described in the previous paragraph is a pod's
ability to use anti-affinity to degrade the scheduling quality of another pod,
but not completely block it from scheduling. For example, a set of pods S1 could
use node affinity to request to schedule onto a set of nodes that some other set
of pods S2 prefers to schedule onto. If the pods in S1 have
RequiredDuringScheduling or even PreferredDuringScheduling pod anti-affinity for
S2, then due to the symmetry property of anti-affinity, they can prevent the
pods in S2 from scheduling onto their preferred nodes if they arrive first (for
sure in the RequiredDuringScheduling case, and with some probability that
depends on the weighting scheme for the PreferredDuringScheduling case). A very
sophisticated priority and/or quota scheme could mitigate this, or alternatively
we could eliminate the symmetry property of the implementation of
PreferredDuringScheduling anti-affinity. Then only RequiredDuringScheduling
anti-affinity could affect scheduling quality of another pod, and as we
described in the previous paragraph, such pods could be charged quota for the
full topology domain, thereby reducing the potential for abuse.
We won't try to address this issue in our initial implementation; we can consider one
of the approaches mentioned above if it turns out to be a problem in practice.
We won't try to address this issue in our initial implementation; we can
consider one of the approaches mentioned above if it turns out to be a problem
in practice.
### Co-existing with daemons
A cluster administrator
may wish to allow pods that express anti-affinity against all pods, to nonetheless co-exist with
system daemon pods, such as those run by DaemonSet. In principle, we would like the specification
for RequiredDuringScheduling inter-pod anti-affinity to allow "toleration" of one or more
other pods (see #18263 for a more detailed explanation of the toleration concept). There are
at least two ways to accomplish this:
A cluster administrator may wish to allow pods that express anti-affinity
against all pods, to nonetheless co-exist with system daemon pods, such as those
run by DaemonSet. In principle, we would like the specification for
RequiredDuringScheduling inter-pod anti-affinity to allow "toleration" of one or
more other pods (see #18263 for a more detailed explanation of the toleration
concept). There are at least two ways to accomplish this:
* Scheduler special-cases the namespace(s) where daemons live, in the
sense that it ignores pods in those namespaces when it is
@ -478,147 +524,168 @@ Our initial implementation will use the first approach.
### Determining which pod(s) to kill (for RequiredDuringSchedulingRequiredDuringExecution)
Because anti-affinity is symmetric, in the case of RequiredDuringSchedulingRequiredDuringExecution
anti-affinity, the system must determine which pod(s) to kill when a pod's labels are updated in
such as way as to cause them to conflict with one or more other pods' RequiredDuringSchedulingRequiredDuringExecution
anti-affinity rules. In the absence of a priority/preemption scheme, our rule will be that the pod
with the anti-affinity rule that becomes violated should be the one killed.
A pod should only specify constraints that apply to
namespaces it trusts to not do malicious things. Once we have priority/preemption, we can
change the rule to say that the lowest-priority pod(s) are killed until all
Because anti-affinity is symmetric, in the case of
RequiredDuringSchedulingRequiredDuringExecution anti-affinity, the system must
determine which pod(s) to kill when a pod's labels are updated in such as way as
to cause them to conflict with one or more other pods'
RequiredDuringSchedulingRequiredDuringExecution anti-affinity rules. In the
absence of a priority/preemption scheme, our rule will be that the pod with the
anti-affinity rule that becomes violated should be the one killed. A pod should
only specify constraints that apply to namespaces it trusts to not do malicious
things. Once we have priority/preemption, we can change the rule to say that the
lowest-priority pod(s) are killed until all
RequiredDuringSchedulingRequiredDuringExecution anti-affinity is satisfied.
## Special considerations for RequiredDuringScheduling affinity
The DoS potential of RequiredDuringScheduling *anti-affinity* stemmed from its symmetry:
if a pod P requests anti-affinity, P cannot schedule onto a node with conflicting pods,
and pods that conflict with P cannot schedule onto the node one P has been scheduled there.
The design we have described says that the symmetry property for RequiredDuringScheduling *affinity*
is weaker: if a pod P says it can only schedule onto nodes running pod Q, this
does not mean Q can only run on a node that is running P, but the scheduler will try
to schedule Q onto a node that is running P (i.e. treats the reverse direction as
preferred). This raises the same scheduling quality concern as we mentioned at the
end of the Denial of Service section above, and can be addressed in similar ways.
The DoS potential of RequiredDuringScheduling *anti-affinity* stemmed from its
symmetry: if a pod P requests anti-affinity, P cannot schedule onto a node with
conflicting pods, and pods that conflict with P cannot schedule onto the node
one P has been scheduled there. The design we have described says that the
symmetry property for RequiredDuringScheduling *affinity* is weaker: if a pod P
says it can only schedule onto nodes running pod Q, this does not mean Q can
only run on a node that is running P, but the scheduler will try to schedule Q
onto a node that is running P (i.e. treats the reverse direction as preferred).
This raises the same scheduling quality concern as we mentioned at the end of
the Denial of Service section above, and can be addressed in similar ways.
The nature of affinity (as opposed to anti-affinity) means that there is no issue of
determining which pod(s) to kill
when a pod's labels change: it is obviously the pod with the affinity rule that becomes
violated that must be killed. (Killing a pod never "fixes" violation of an affinity rule;
it can only "fix" violation an anti-affinity rule.) However, affinity does have a
different question related to killing: how long should the system wait before declaring
that RequiredDuringSchedulingRequiredDuringExecution affinity is no longer met at runtime?
For example, if a pod P has such an affinity for a pod Q and pod Q is temporarily killed
so that it can be updated to a new binary version, should that trigger killing of P? More
generally, how long should the system wait before declaring that P's affinity is
violated? (Of course affinity is expressed in terms of label selectors, not for a specific
pod, but the scenario is easier to describe using a concrete pod.) This is closely related to
the concept of forgiveness (see issue #1574). In theory we could make this time duration be
configurable by the user on a per-pod basis, but for the first version of this feature we will
make it a configurable property of whichever component does the killing and that applies across
all pods using the feature. Making it configurable by the user would require a nontrivial change
to the API syntax (since the field would only apply to RequiredDuringSchedulingRequiredDuringExecution
affinity).
The nature of affinity (as opposed to anti-affinity) means that there is no
issue of determining which pod(s) to kill when a pod's labels change: it is
obviously the pod with the affinity rule that becomes violated that must be
killed. (Killing a pod never "fixes" violation of an affinity rule; it can only
"fix" violation an anti-affinity rule.) However, affinity does have a different
question related to killing: how long should the system wait before declaring
that RequiredDuringSchedulingRequiredDuringExecution affinity is no longer met
at runtime? For example, if a pod P has such an affinity for a pod Q and pod Q
is temporarily killed so that it can be updated to a new binary version, should
that trigger killing of P? More generally, how long should the system wait
before declaring that P's affinity is violated? (Of course affinity is expressed
in terms of label selectors, not for a specific pod, but the scenario is easier
to describe using a concrete pod.) This is closely related to the concept of
forgiveness (see issue #1574). In theory we could make this time duration be
configurable by the user on a per-pod basis, but for the first version of this
feature we will make it a configurable property of whichever component does the
killing and that applies across all pods using the feature. Making it
configurable by the user would require a nontrivial change to the API syntax
(since the field would only apply to
RequiredDuringSchedulingRequiredDuringExecution affinity).
## Implementation plan
1. Add the `Affinity` field to PodSpec and the `PodAffinity` and `PodAntiAffinity` types to the API along with all of their descendant types.
2. Implement a scheduler predicate that takes `RequiredDuringSchedulingIgnoredDuringExecution`
affinity and anti-affinity into account. Include a workaround for the issue described at the end of the Affinity section of the Examples section (can't schedule first pod).
3. Implement a scheduler priority function that takes `PreferredDuringSchedulingIgnoredDuringExecution` affinity and anti-affinity into account
4. Implement admission controller that rejects requests that specify "all namespaces" with non-"node" TopologyKey for `RequiredDuringScheduling` anti-affinity.
This admission controller should be enabled by default.
1. Add the `Affinity` field to PodSpec and the `PodAffinity` and
`PodAntiAffinity` types to the API along with all of their descendant types.
2. Implement a scheduler predicate that takes
`RequiredDuringSchedulingIgnoredDuringExecution` affinity and anti-affinity into
account. Include a workaround for the issue described at the end of the Affinity
section of the Examples section (can't schedule first pod).
3. Implement a scheduler priority function that takes
`PreferredDuringSchedulingIgnoredDuringExecution` affinity and anti-affinity
into account.
4. Implement admission controller that rejects requests that specify "all
namespaces" with non-"node" TopologyKey for `RequiredDuringScheduling`
anti-affinity. This admission controller should be enabled by default.
5. Implement the recommended solution to the "co-existing with daemons" issue
6. At this point, the feature can be deployed.
7. Add the `RequiredDuringSchedulingRequiredDuringExecution` field to affinity and anti-affinity, and make sure
the pieces of the system already implemented for `RequiredDuringSchedulingIgnoredDuringExecution` also take
`RequiredDuringSchedulingRequiredDuringExecution` into account (e.g. the scheduler predicate, the quota mechanism,
the "co-existing with daemons" solution).
8. Add `RequiredDuringSchedulingRequiredDuringExecution` for "node" `TopologyKey` to Kubelet's admission decision
9. Implement code in Kubelet *or* the controllers that evicts a pod that no longer satisfies
`RequiredDuringSchedulingRequiredDuringExecution`. If Kubelet then only for "node" `TopologyKey`;
if controller then potentially for all `TopologyKeys`'s.
(see [this comment](https://github.com/kubernetes/kubernetes/issues/12744#issuecomment-164372008)).
7. Add the `RequiredDuringSchedulingRequiredDuringExecution` field to affinity
and anti-affinity, and make sure the pieces of the system already implemented
for `RequiredDuringSchedulingIgnoredDuringExecution` also take
`RequiredDuringSchedulingRequiredDuringExecution` into account (e.g. the
scheduler predicate, the quota mechanism, the "co-existing with daemons"
solution).
8. Add `RequiredDuringSchedulingRequiredDuringExecution` for "node"
`TopologyKey` to Kubelet's admission decision.
9. Implement code in Kubelet *or* the controllers that evicts a pod that no
longer satisfies `RequiredDuringSchedulingRequiredDuringExecution`. If Kubelet
then only for "node" `TopologyKey`; if controller then potentially for all
`TopologyKeys`'s. (see [this comment](https://github.com/kubernetes/kubernetes/issues/12744#issuecomment-164372008)).
Do so in a way that addresses the "determining which pod(s) to kill" issue.
We assume Kubelet publishes labels describing the node's membership in all of the relevant scheduling
domains (e.g. node name, rack name, availability zone name, etc.). See #9044.
We assume Kubelet publishes labels describing the node's membership in all of
the relevant scheduling domains (e.g. node name, rack name, availability zone
name, etc.). See #9044.
## Backward compatibility
Old versions of the scheduler will ignore `Affinity`.
Users should not start using `Affinity` until the full implementation has
been in Kubelet and the master for enough binary versions that we feel
comfortable that we will not need to roll back either Kubelet or
master to a version that does not support them. Longer-term we will
use a programmatic approach to enforcing this (#4855).
Users should not start using `Affinity` until the full implementation has been
in Kubelet and the master for enough binary versions that we feel comfortable
that we will not need to roll back either Kubelet or master to a version that
does not support them. Longer-term we will use a programmatic approach to
enforcing this (#4855).
## Extensibility
The design described here is the result of careful analysis of use cases, a decade of experience
with Borg at Google, and a review of similar features in other open-source container orchestration
systems. We believe that it properly balances the goal of expressiveness against the goals of
simplicity and efficiency of implementation. However, we recognize that
use cases may arise in the future that cannot be expressed using the syntax described here.
Although we are not implementing an affinity-specific extensibility mechanism for a variety
of reasons (simplicity of the codebase, simplicity of cluster deployment, desire for Kubernetes
users to get a consistent experience, etc.), the regular Kubernetes
annotation mechanism can be used to add or replace affinity rules. The way this work would is
The design described here is the result of careful analysis of use cases, a
decade of experience with Borg at Google, and a review of similar features in
other open-source container orchestration systems. We believe that it properly
balances the goal of expressiveness against the goals of simplicity and
efficiency of implementation. However, we recognize that use cases may arise in
the future that cannot be expressed using the syntax described here. Although we
are not implementing an affinity-specific extensibility mechanism for a variety
of reasons (simplicity of the codebase, simplicity of cluster deployment, desire
for Kubernetes users to get a consistent experience, etc.), the regular
Kubernetes annotation mechanism can be used to add or replace affinity rules.
The way this work would is:
1. Define one or more annotations to describe the new affinity rule(s)
1. User (or an admission controller) attaches the annotation(s) to pods to request the desired scheduling behavior.
If the new rule(s) *replace* one or more fields of `Affinity` then the user would omit those fields
from `Affinity`; if they are *additional rules*, then the user would fill in `Affinity` as well as the
annotation(s).
1. User (or an admission controller) attaches the annotation(s) to pods to
request the desired scheduling behavior. If the new rule(s) *replace* one or
more fields of `Affinity` then the user would omit those fields from `Affinity`;
if they are *additional rules*, then the user would fill in `Affinity` as well
as the annotation(s).
1. Scheduler takes the annotation(s) into account when scheduling.
If some particular new syntax becomes popular, we would consider upstreaming it by integrating
it into the standard `Affinity`.
If some particular new syntax becomes popular, we would consider upstreaming it
by integrating it into the standard `Affinity`.
## Future work and non-work
One can imagine that in the anti-affinity RequiredDuringScheduling case
one might want to associate a number with the rule,
for example "do not allow this pod to share a rack with more than three other
pods (in total, or from the same service as the pod)." We could allow this to be
specified by adding an integer `Limit` to `PodAffinityTerm` just for the
`RequiredDuringScheduling` case. However, this flexibility complicates the
system and we do not intend to implement it.
One can imagine that in the anti-affinity RequiredDuringScheduling case one
might want to associate a number with the rule, for example "do not allow this
pod to share a rack with more than three other pods (in total, or from the same
service as the pod)." We could allow this to be specified by adding an integer
`Limit` to `PodAffinityTerm` just for the `RequiredDuringScheduling` case.
However, this flexibility complicates the system and we do not intend to
implement it.
It is likely that the specification and implementation of pod anti-affinity
can be unified with [taints and tolerations](taint-toleration-dedicated.md),
and likewise that the specification and implementation of pod affinity
can be unified with [node affinity](nodeaffinity.md).
The basic idea is that pod labels would be "inherited" by the node, and pods
would only be able to specify affinity and anti-affinity for a node's labels.
Our main motivation for not unifying taints and tolerations with
pod anti-affinity is that we foresee taints and tolerations as being a concept that
only cluster administrators need to understand (and indeed in some setups taints and
tolerations wouldn't even be directly manipulated by a cluster administrator,
instead they would only be set by an admission controller that is implementing the administrator's
high-level policy about different classes of special machines and the users who belong to the groups
allowed to access them). Moreover, the concept of nodes "inheriting" labels
from pods seems complicated; it seems conceptually simpler to separate rules involving
relatively static properties of nodes from rules involving which other pods are running
on the same node or larger topology domain.
can be unified with [node affinity](nodeaffinity.md). The basic idea is that pod
labels would be "inherited" by the node, and pods would only be able to specify
affinity and anti-affinity for a node's labels. Our main motivation for not
unifying taints and tolerations with pod anti-affinity is that we foresee taints
and tolerations as being a concept that only cluster administrators need to
understand (and indeed in some setups taints and tolerations wouldn't even be
directly manipulated by a cluster administrator, instead they would only be set
by an admission controller that is implementing the administrator's high-level
policy about different classes of special machines and the users who belong to
the groups allowed to access them). Moreover, the concept of nodes "inheriting"
labels from pods seems complicated; it seems conceptually simpler to separate
rules involving relatively static properties of nodes from rules involving which
other pods are running on the same node or larger topology domain.
Data/storage affinity is related to pod affinity, and is likely to draw on some of the
ideas we have used for pod affinity. Today, data/storage affinity is expressed using
node affinity, on the assumption that the pod knows which node(s) store(s) the data
it wants. But a more flexible approach would allow the pod to name the data rather than
the node.
Data/storage affinity is related to pod affinity, and is likely to draw on some
of the ideas we have used for pod affinity. Today, data/storage affinity is
expressed using node affinity, on the assumption that the pod knows which
node(s) store(s) the data it wants. But a more flexible approach would allow the
pod to name the data rather than the node.
## Related issues
The review for this proposal is in #18265.
The topic of affinity/anti-affinity has generated a lot of discussion. The main issue
is #367 but #14484/#14485, #9560, #11369, #14543, #11707, #3945, #341, #1965, and #2906
all have additional discussion and use cases.
The topic of affinity/anti-affinity has generated a lot of discussion. The main
issue is #367 but #14484/#14485, #9560, #11369, #14543, #11707, #3945, #341,
As the examples in this document have demonstrated, topological affinity is very useful
in clusters that are spread across availability zones, e.g. to co-locate pods of a service
in the same zone to avoid a wide-area network hop, or to spread pods across zones for
failure tolerance. #17059, #13056, #13063, and #4235 are relevant.
# 1965, and #2906 all have additional discussion and use cases.
As the examples in this document have demonstrated, topological affinity is very
useful in clusters that are spread across availability zones, e.g. to co-locate
pods of a service in the same zone to avoid a wide-area network hop, or to
spread pods across zones for failure tolerance. #17059, #13056, #13063, and
# 4235 are relevant.
Issue #15675 describes connection affinity, which is vaguely related.

77
docs/design/principles.md
View File

@ -43,26 +43,57 @@ See also the [API conventions](../devel/api-conventions.md).
* All APIs should be declarative.
* API objects should be complementary and composable, not opaque wrappers.
* The control plane should be transparent -- there are no hidden internal APIs.
* The cost of API operations should be proportional to the number of objects intentionally operated upon. Therefore, common filtered lookups must be indexed. Beware of patterns of multiple API calls that would incur quadratic behavior.
* Object status must be 100% reconstructable by observation. Any history kept must be just an optimization and not required for correct operation.
* Cluster-wide invariants are difficult to enforce correctly. Try not to add them. If you must have them, don't enforce them atomically in master components, that is contention-prone and doesn't provide a recovery path in the case of a bug allowing the invariant to be violated. Instead, provide a series of checks to reduce the probability of a violation, and make every component involved able to recover from an invariant violation.
* Low-level APIs should be designed for control by higher-level systems. Higher-level APIs should be intent-oriented (think SLOs) rather than implementation-oriented (think control knobs).
* The cost of API operations should be proportional to the number of objects
intentionally operated upon. Therefore, common filtered lookups must be indexed.
Beware of patterns of multiple API calls that would incur quadratic behavior.
* Object status must be 100% reconstructable by observation. Any history kept
must be just an optimization and not required for correct operation.
* Cluster-wide invariants are difficult to enforce correctly. Try not to add
them. If you must have them, don't enforce them atomically in master components,
that is contention-prone and doesn't provide a recovery path in the case of a
bug allowing the invariant to be violated. Instead, provide a series of checks
to reduce the probability of a violation, and make every component involved able
to recover from an invariant violation.
* Low-level APIs should be designed for control by higher-level systems.
Higher-level APIs should be intent-oriented (think SLOs) rather than
implementation-oriented (think control knobs).
## Control logic
* Functionality must be *level-based*, meaning the system must operate correctly given the desired state and the current/observed state, regardless of how many intermediate state updates may have been missed. Edge-triggered behavior must be just an optimization.
* Assume an open world: continually verify assumptions and gracefully adapt to external events and/or actors. Example: we allow users to kill pods under control of a replication controller; it just replaces them.
* Do not define comprehensive state machines for objects with behaviors associated with state transitions and/or "assumed" states that cannot be ascertained by observation.
* Don't assume a component's decisions will not be overridden or rejected, nor for the component to always understand why. For example, etcd may reject writes. Kubelet may reject pods. The scheduler may not be able to schedule pods. Retry, but back off and/or make alternative decisions.
* Components should be self-healing. For example, if you must keep some state (e.g., cache) the content needs to be periodically refreshed, so that if an item does get erroneously stored or a deletion event is missed etc, it will be soon fixed, ideally on timescales that are shorter than what will attract attention from humans.
* Component behavior should degrade gracefully. Prioritize actions so that the most important activities can continue to function even when overloaded and/or in states of partial failure.
* Functionality must be *level-based*, meaning the system must operate correctly
given the desired state and the current/observed state, regardless of how many
intermediate state updates may have been missed. Edge-triggered behavior must be
just an optimization.
* Assume an open world: continually verify assumptions and gracefully adapt to
external events and/or actors. Example: we allow users to kill pods under
control of a replication controller; it just replaces them.
* Do not define comprehensive state machines for objects with behaviors
associated with state transitions and/or "assumed" states that cannot be
ascertained by observation.
* Don't assume a component's decisions will not be overridden or rejected, nor
for the component to always understand why. For example, etcd may reject writes.
Kubelet may reject pods. The scheduler may not be able to schedule pods. Retry,
but back off and/or make alternative decisions.
* Components should be self-healing. For example, if you must keep some state
(e.g., cache) the content needs to be periodically refreshed, so that if an item
does get erroneously stored or a deletion event is missed etc, it will be soon
fixed, ideally on timescales that are shorter than what will attract attention
from humans.
* Component behavior should degrade gracefully. Prioritize actions so that the
most important activities can continue to function even when overloaded and/or
in states of partial failure.
## Architecture
* Only the apiserver should communicate with etcd/store, and not other components (scheduler, kubelet, etc.).
* Only the apiserver should communicate with etcd/store, and not other
components (scheduler, kubelet, etc.).
* Compromising a single node shouldn't compromise the cluster.
* Components should continue to do what they were last told in the absence of new instructions (e.g., due to network partition or component outage).
* All components should keep all relevant state in memory all the time. The apiserver should write through to etcd/store, other components should write through to the apiserver, and they should watch for updates made by other clients.
* Components should continue to do what they were last told in the absence of
new instructions (e.g., due to network partition or component outage).
* All components should keep all relevant state in memory all the time. The
apiserver should write through to etcd/store, other components should write
through to the apiserver, and they should watch for updates made by other
clients.
* Watch is preferred over polling.
## Extensibility
@ -72,13 +103,23 @@ TODO: pluggability
## Bootstrapping
* [Self-hosting](http://issue.k8s.io/246) of all components is a goal.
* Minimize the number of dependencies, particularly those required for steady-state operation.
* Minimize the number of dependencies, particularly those required for
steady-state operation.
* Stratify the dependencies that remain via principled layering.
* Break any circular dependencies by converting hard dependencies to soft dependencies.
* Also accept that data from other components from another source, such as local files, which can then be manually populated at bootstrap time and then continuously updated once those other components are available.
* Break any circular dependencies by converting hard dependencies to soft
dependencies.
* Also accept that data from other components from another source, such as
local files, which can then be manually populated at bootstrap time and then
continuously updated once those other components are available.
* State should be rediscoverable and/or reconstructable.
* Make it easy to run temporary, bootstrap instances of all components in order to create the runtime state needed to run the components in the steady state; use a lock (master election for distributed components, file lock for local components like Kubelet) to coordinate handoff. We call this technique "pivoting".
* Have a solution to restart dead components. For distributed components, replication works well. For local components such as Kubelet, a process manager or even a simple shell loop works.
* Make it easy to run temporary, bootstrap instances of all components in
order to create the runtime state needed to run the components in the steady
state; use a lock (master election for distributed components, file lock for
local components like Kubelet) to coordinate handoff. We call this technique
"pivoting".
* Have a solution to restart dead components. For distributed components,
replication works well. For local components such as Kubelet, a process manager
or even a simple shell loop works.
## Availability

237
docs/design/resources.md
View File

@ -31,16 +31,19 @@ Documentation for other releases can be found at
<!-- END STRIP_FOR_RELEASE -->
<!-- END MUNGE: UNVERSIONED_WARNING -->
**Note: this is a design doc, which describes features that have not been completely implemented.
User documentation of the current state is [here](../user-guide/compute-resources.md). The tracking issue for
implementation of this model is
[#168](http://issue.k8s.io/168). Currently, both limits and requests of memory and
cpu on containers (not pods) are supported. "memory" is in bytes and "cpu" is in
milli-cores.**
**Note: this is a design doc, which describes features that have not been
completely implemented. User documentation of the current state is
[here](../user-guide/compute-resources.md). The tracking issue for
implementation of this model is [#168](http://issue.k8s.io/168). Currently, both
limits and requests of memory and cpu on containers (not pods) are supported.
"memory" is in bytes and "cpu" is in milli-cores.**
# The Kubernetes resource model
To do good pod placement, Kubernetes needs to know how big pods are, as well as the sizes of the nodes onto which they are being placed. The definition of "how big" is given by the Kubernetes resource model &mdash; the subject of this document.
To do good pod placement, Kubernetes needs to know how big pods are, as well as
the sizes of the nodes onto which they are being placed. The definition of "how
big" is given by the Kubernetes resource model &mdash; the subject of this
document.
The resource model aims to be:
* simple, for common cases;
@ -50,43 +53,107 @@ The resource model aims to be:
## The resource model
A Kubernetes _resource_ is something that can be requested by, allocated to, or consumed by a pod or container. Examples include memory (RAM), CPU, disk-time, and network bandwidth.
A Kubernetes _resource_ is something that can be requested by, allocated to, or
consumed by a pod or container. Examples include memory (RAM), CPU, disk-time,
and network bandwidth.
Once resources on a node have been allocated to one pod, they should not be allocated to another until that pod is removed or exits. This means that Kubernetes schedulers should ensure that the sum of the resources allocated (requested and granted) to its pods never exceeds the usable capacity of the node. Testing whether a pod will fit on a node is called _feasibility checking_.
Once resources on a node have been allocated to one pod, they should not be
allocated to another until that pod is removed or exits. This means that
Kubernetes schedulers should ensure that the sum of the resources allocated
(requested and granted) to its pods never exceeds the usable capacity of the
node. Testing whether a pod will fit on a node is called _feasibility checking_.
Note that the resource model currently prohibits over-committing resources; we will want to relax that restriction later.
Note that the resource model currently prohibits over-committing resources; we
will want to relax that restriction later.
### Resource types
All resources have a _type_ that is identified by their _typename_ (a string, e.g., "memory"). Several resource types are predefined by Kubernetes (a full list is below), although only two will be supported at first: CPU and memory. Users and system administrators can define their own resource types if they wish (e.g., Hadoop slots).
All resources have a _type_ that is identified by their _typename_ (a string,
e.g., "memory"). Several resource types are predefined by Kubernetes (a full
list is below), although only two will be supported at first: CPU and memory.
Users and system administrators can define their own resource types if they wish
(e.g., Hadoop slots).
A fully-qualified resource typename is constructed from a DNS-style _subdomain_, followed by a slash `/`, followed by a name.
* The subdomain must conform to [RFC 1123](http://www.ietf.org/rfc/rfc1123.txt) (e.g., `kubernetes.io`, `example.com`).
* The name must be not more than 63 characters, consisting of upper- or lower-case alphanumeric characters, with the `-`, `_`, and `.` characters allowed anywhere except the first or last character.
* As a shorthand, any resource typename that does not start with a subdomain and a slash will automatically be prefixed with the built-in Kubernetes _namespace_, `kubernetes.io/` in order to fully-qualify it. This namespace is reserved for code in the open source Kubernetes repository; as a result, all user typenames MUST be fully qualified, and cannot be created in this namespace.
A fully-qualified resource typename is constructed from a DNS-style _subdomain_,
followed by a slash `/`, followed by a name.
* The subdomain must conform to [RFC 1123](http://www.ietf.org/rfc/rfc1123.txt)
(e.g., `kubernetes.io`, `example.com`).
* The name must be not more than 63 characters, consisting of upper- or
lower-case alphanumeric characters, with the `-`, `_`, and `.` characters
allowed anywhere except the first or last character.
* As a shorthand, any resource typename that does not start with a subdomain and
a slash will automatically be prefixed with the built-in Kubernetes _namespace_,
`kubernetes.io/` in order to fully-qualify it. This namespace is reserved for
code in the open source Kubernetes repository; as a result, all user typenames
MUST be fully qualified, and cannot be created in this namespace.
Some example typenames include `memory` (which will be fully-qualified as `kubernetes.io/memory`), and `example.com/Shiny_New-Resource.Type`.
Some example typenames include `memory` (which will be fully-qualified as
`kubernetes.io/memory`), and `example.com/Shiny_New-Resource.Type`.
For future reference, note that some resources, such as CPU and network bandwidth, are _compressible_, which means that their usage can potentially be throttled in a relatively benign manner. All other resources are _incompressible_, which means that any attempt to throttle them is likely to cause grief. This distinction will be important if a Kubernetes implementation supports over-committing of resources.
For future reference, note that some resources, such as CPU and network
bandwidth, are _compressible_, which means that their usage can potentially be
throttled in a relatively benign manner. All other resources are
_incompressible_, which means that any attempt to throttle them is likely to
cause grief. This distinction will be important if a Kubernetes implementation
supports over-committing of resources.
### Resource quantities
Initially, all Kubernetes resource types are _quantitative_, and have an associated _unit_ for quantities of the associated resource (e.g., bytes for memory, bytes per seconds for bandwidth, instances for software licences). The units will always be a resource type's natural base units (e.g., bytes, not MB), to avoid confusion between binary and decimal multipliers and the underlying unit multiplier (e.g., is memory measured in MiB, MB, or GB?).
Initially, all Kubernetes resource types are _quantitative_, and have an
associated _unit_ for quantities of the associated resource (e.g., bytes for
memory, bytes per seconds for bandwidth, instances for software licences). The
units will always be a resource type's natural base units (e.g., bytes, not MB),
to avoid confusion between binary and decimal multipliers and the underlying
unit multiplier (e.g., is memory measured in MiB, MB, or GB?).
Resource quantities can be added and subtracted: for example, a node has a fixed quantity of each resource type that can be allocated to pods/containers; once such an allocation has been made, the allocated resources cannot be made available to other pods/containers without over-committing the resources.
Resource quantities can be added and subtracted: for example, a node has a fixed
quantity of each resource type that can be allocated to pods/containers; once
such an allocation has been made, the allocated resources cannot be made
available to other pods/containers without over-committing the resources.
To make life easier for people, quantities can be represented externally as unadorned integers, or as fixed-point integers with one of these SI suffices (E, P, T, G, M, K, m) or their power-of-two equivalents (Ei, Pi, Ti, Gi, Mi, Ki). For example, the following represent roughly the same value: 128974848, "129e6", "129M" , "123Mi". Small quantities can be represented directly as decimals (e.g., 0.3), or using milli-units (e.g., "300m").
* "Externally" means in user interfaces, reports, graphs, and in JSON or YAML resource specifications that might be generated or read by people.
* Case is significant: "m" and "M" are not the same, so "k" is not a valid SI suffix. There are no power-of-two equivalents for SI suffixes that represent multipliers less than 1.
To make life easier for people, quantities can be represented externally as
unadorned integers, or as fixed-point integers with one of these SI suffices
(E, P, T, G, M, K, m) or their power-of-two equivalents (Ei, Pi, Ti, Gi, Mi,
Ki). For example, the following represent roughly the same value: 128974848,
"129e6", "129M" , "123Mi". Small quantities can be represented directly as
decimals (e.g., 0.3), or using milli-units (e.g., "300m").
* "Externally" means in user interfaces, reports, graphs, and in JSON or YAML
resource specifications that might be generated or read by people.
* Case is significant: "m" and "M" are not the same, so "k" is not a valid SI
suffix. There are no power-of-two equivalents for SI suffixes that represent
multipliers less than 1.
* These conventions only apply to resource quantities, not arbitrary values.
Internally (i.e., everywhere else), Kubernetes will represent resource quantities as integers so it can avoid problems with rounding errors, and will not use strings to represent numeric values. To achieve this, quantities that naturally have fractional parts (e.g., CPU seconds/second) will be scaled to integral numbers of milli-units (e.g., milli-CPUs) as soon as they are read in. Internal APIs, data structures, and protobufs will use these scaled integer units. Raw measurement data such as usage may still need to be tracked and calculated using floating point values, but internally they should be rescaled to avoid some values being in milli-units and some not.
* Note that reading in a resource quantity and writing it out again may change the way its values are represented, and truncate precision (e.g., 1.0001 may become 1.000), so comparison and difference operations (e.g., by an updater) must be done on the internal representations.
* Avoiding milli-units in external representations has advantages for people who will use Kubernetes, but runs the risk of developers forgetting to rescale or accidentally using floating-point representations. That seems like the right choice. We will try to reduce the risk by providing libraries that automatically do the quantization for JSON/YAML inputs.
Internally (i.e., everywhere else), Kubernetes will represent resource
quantities as integers so it can avoid problems with rounding errors, and will
not use strings to represent numeric values. To achieve this, quantities that
naturally have fractional parts (e.g., CPU seconds/second) will be scaled to
integral numbers of milli-units (e.g., milli-CPUs) as soon as they are read in.
Internal APIs, data structures, and protobufs will use these scaled integer
units. Raw measurement data such as usage may still need to be tracked and
calculated using floating point values, but internally they should be rescaled
to avoid some values being in milli-units and some not.
* Note that reading in a resource quantity and writing it out again may change
the way its values are represented, and truncate precision (e.g., 1.0001 may
become 1.000), so comparison and difference operations (e.g., by an updater)
must be done on the internal representations.
* Avoiding milli-units in external representations has advantages for people
who will use Kubernetes, but runs the risk of developers forgetting to rescale
or accidentally using floating-point representations. That seems like the right
choice. We will try to reduce the risk by providing libraries that automatically
do the quantization for JSON/YAML inputs.
### Resource specifications
Both users and a number of system components, such as schedulers, (horizontal) auto-scalers, (vertical) auto-sizers, load balancers, and worker-pool managers need to reason about resource requirements of workloads, resource capacities of nodes, and resource usage. Kubernetes divides specifications of *desired state*, aka the Spec, and representations of *current state*, aka the Status. Resource requirements and total node capacity fall into the specification category, while resource usage, characterizations derived from usage (e.g., maximum usage, histograms), and other resource demand signals (e.g., CPU load) clearly fall into the status category and are discussed in the Appendix for now.
Both users and a number of system components, such as schedulers, (horizontal)
auto-scalers, (vertical) auto-sizers, load balancers, and worker-pool managers
need to reason about resource requirements of workloads, resource capacities of
nodes, and resource usage. Kubernetes divides specifications of *desired state*,
aka the Spec, and representations of *current state*, aka the Status. Resource
requirements and total node capacity fall into the specification category, while
resource usage, characterizations derived from usage (e.g., maximum usage,
histograms), and other resource demand signals (e.g., CPU load) clearly fall
into the status category and are discussed in the Appendix for now.
Resource requirements for a container or pod should have the following form:
@ -98,9 +165,24 @@ resourceRequirementSpec: [
```
Where:
* _request_ [optional]: the amount of resources being requested, or that were requested and have been allocated. Scheduler algorithms will use these quantities to test feasibility (whether a pod will fit onto a node). If a container (or pod) tries to use more resources than its _request_, any associated SLOs are voided &mdash; e.g., the program it is running may be throttled (compressible resource types), or the attempt may be denied. If _request_ is omitted for a container, it defaults to _limit_ if that is explicitly specified, otherwise to an implementation-defined value; this will always be 0 for a user-defined resource type. If _request_ is omitted for a pod, it defaults to the sum of the (explicit or implicit) _request_ values for the containers it encloses.
* _request_ [optional]: the amount of resources being requested, or that were
requested and have been allocated. Scheduler algorithms will use these
quantities to test feasibility (whether a pod will fit onto a node).
If a container (or pod) tries to use more resources than its _request_, any
associated SLOs are voided &mdash; e.g., the program it is running may be
throttled (compressible resource types), or the attempt may be denied. If
_request_ is omitted for a container, it defaults to _limit_ if that is
explicitly specified, otherwise to an implementation-defined value; this will
always be 0 for a user-defined resource type. If _request_ is omitted for a pod,
it defaults to the sum of the (explicit or implicit) _request_ values for the
containers it encloses.
* _limit_ [optional]: an upper bound or cap on the maximum amount of resources that will be made available to a container or pod; if a container or pod uses more resources than its _limit_, it may be terminated. The _limit_ defaults to "unbounded"; in practice, this probably means the capacity of an enclosing container, pod, or node, but may result in non-deterministic behavior, especially for memory.
* _limit_ [optional]: an upper bound or cap on the maximum amount of resources
that will be made available to a container or pod; if a container or pod uses
more resources than its _limit_, it may be terminated. The _limit_ defaults to
"unbounded"; in practice, this probably means the capacity of an enclosing
container, pod, or node, but may result in non-deterministic behavior,
especially for memory.
Total capacity for a node should have a similar structure:
@ -111,36 +193,66 @@ resourceCapacitySpec: [
```
Where:
* _total_: the total allocatable resources of a node. Initially, the resources at a given scope will bound the resources of the sum of inner scopes.
* _total_: the total allocatable resources of a node. Initially, the resources
at a given scope will bound the resources of the sum of inner scopes.
#### Notes
* It is an error to specify the same resource type more than once in each list.
* It is an error to specify the same resource type more than once in each
list.
* It is an error for the _request_ or _limit_ values for a pod to be less than the sum of the (explicit or defaulted) values for the containers it encloses. (We may relax this later.)
* It is an error for the _request_ or _limit_ values for a pod to be less than
the sum of the (explicit or defaulted) values for the containers it encloses.
(We may relax this later.)
* If multiple pods are running on the same node and attempting to use more resources than they have requested, the result is implementation-defined. For example: unallocated or unused resources might be spread equally across claimants, or the assignment might be weighted by the size of the original request, or as a function of limits, or priority, or the phase of the moon, perhaps modulated by the direction of the tide. Thus, although it's not mandatory to provide a _request_, it's probably a good idea. (Note that the _request_ could be filled in by an automated system that is observing actual usage and/or historical data.)
* If multiple pods are running on the same node and attempting to use more
resources than they have requested, the result is implementation-defined. For
example: unallocated or unused resources might be spread equally across
claimants, or the assignment might be weighted by the size of the original
request, or as a function of limits, or priority, or the phase of the moon,
perhaps modulated by the direction of the tide. Thus, although it's not
mandatory to provide a _request_, it's probably a good idea. (Note that the
_request_ could be filled in by an automated system that is observing actual
usage and/or historical data.)
* Internally, the Kubernetes master can decide the defaulting behavior and the kubelet implementation may expected an absolute specification. For example, if the master decided that "the default is unbounded" it would pass 2^64 to the kubelet.
* Internally, the Kubernetes master can decide the defaulting behavior and the
kubelet implementation may expected an absolute specification. For example, if
the master decided that "the default is unbounded" it would pass 2^64 to the
kubelet.
## Kubernetes-defined resource types
The following resource types are predefined ("reserved") by Kubernetes in the `kubernetes.io` namespace, and so cannot be used for user-defined resources. Note that the syntax of all resource types in the resource spec is deliberately similar, but some resource types (e.g., CPU) may receive significantly more support than simply tracking quantities in the schedulers and/or the Kubelet.
The following resource types are predefined ("reserved") by Kubernetes in the
`kubernetes.io` namespace, and so cannot be used for user-defined resources.
Note that the syntax of all resource types in the resource spec is deliberately
similar, but some resource types (e.g., CPU) may receive significantly more
support than simply tracking quantities in the schedulers and/or the Kubelet.
### Processor cycles
* Name: `cpu` (or `kubernetes.io/cpu`)
* Units: Kubernetes Compute Unit seconds/second (i.e., CPU cores normalized to a canonical "Kubernetes CPU")
* Units: Kubernetes Compute Unit seconds/second (i.e., CPU cores normalized to
a canonical "Kubernetes CPU")
* Internal representation: milli-KCUs
* Compressible? yes
* Qualities: this is a placeholder for the kind of thing that may be supported in the future &mdash; see [#147](http://issue.k8s.io/147)
* Qualities: this is a placeholder for the kind of thing that may be supported
in the future &mdash; see [#147](http://issue.k8s.io/147)
* [future] `schedulingLatency`: as per lmctfy
* [future] `cpuConversionFactor`: property of a node: the speed of a CPU core on the node's processor divided by the speed of the canonical Kubernetes CPU (a floating point value; default = 1.0).
* [future] `cpuConversionFactor`: property of a node: the speed of a CPU
core on the node's processor divided by the speed of the canonical Kubernetes
CPU (a floating point value; default = 1.0).
To reduce performance portability problems for pods, and to avoid worse-case provisioning behavior, the units of CPU will be normalized to a canonical "Kubernetes Compute Unit" (KCU, pronounced ˈko͝oko͞o), which will roughly be equivalent to a single CPU hyperthreaded core for some recent x86 processor. The normalization may be implementation-defined, although some reasonable defaults will be provided in the open-source Kubernetes code.
To reduce performance portability problems for pods, and to avoid worse-case
provisioning behavior, the units of CPU will be normalized to a canonical
"Kubernetes Compute Unit" (KCU, pronounced ˈko͝oko͞o), which will roughly be
equivalent to a single CPU hyperthreaded core for some recent x86 processor. The
normalization may be implementation-defined, although some reasonable defaults
will be provided in the open-source Kubernetes code.
Note that requesting 2 KCU won't guarantee that precisely 2 physical cores will be allocated &mdash; control of aspects like this will be handled by resource _qualities_ (a future feature).
Note that requesting 2 KCU won't guarantee that precisely 2 physical cores will
be allocated &mdash; control of aspects like this will be handled by resource
_qualities_ (a future feature).
### Memory
@ -149,15 +261,18 @@ Note that requesting 2 KCU won't guarantee that precisely 2 physical cores will
* Units: bytes
* Compressible? no (at least initially)
The precise meaning of what "memory" means is implementation dependent, but the basic idea is to rely on the underlying `memcg` mechanisms, support, and definitions.
The precise meaning of what "memory" means is implementation dependent, but the
basic idea is to rely on the underlying `memcg` mechanisms, support, and
definitions.
Note that most people will want to use power-of-two suffixes (Mi, Gi) for memory quantities
rather than decimal ones: "64MiB" rather than "64MB".
Note that most people will want to use power-of-two suffixes (Mi, Gi) for memory
quantities rather than decimal ones: "64MiB" rather than "64MB".
## Resource metadata
A resource type may have an associated read-only ResourceType structure, that contains metadata about the type. For example:
A resource type may have an associated read-only ResourceType structure, that
contains metadata about the type. For example:
```yaml
resourceTypes: [
@ -172,7 +287,10 @@ resourceTypes: [
]
```
Kubernetes will provide ResourceType metadata for its predefined types. If no resource metadata can be found for a resource type, Kubernetes will assume that it is a quantified, incompressible resource that is not specified in milli-units, and has no default value.
Kubernetes will provide ResourceType metadata for its predefined types. If no
resource metadata can be found for a resource type, Kubernetes will assume that
it is a quantified, incompressible resource that is not specified in
milli-units, and has no default value.
The defined properties are as follows:
@ -188,13 +306,21 @@ The defined properties are as follows:
# Appendix: future extensions
The following are planned future extensions to the resource model, included here to encourage comments.
The following are planned future extensions to the resource model, included here
to encourage comments.
## Usage data
Because resource usage and related metrics change continuously, need to be tracked over time (i.e., historically), can be characterized in a variety of ways, and are fairly voluminous, we will not include usage in core API objects, such as [Pods](../user-guide/pods.md) and Nodes, but will provide separate APIs for accessing and managing that data. See the Appendix for possible representations of usage data, but the representation we'll use is TBD.
Because resource usage and related metrics change continuously, need to be
tracked over time (i.e., historically), can be characterized in a variety of
ways, and are fairly voluminous, we will not include usage in core API objects,
such as [Pods](../user-guide/pods.md) and Nodes, but will provide separate APIs
for accessing and managing that data. See the Appendix for possible
representations of usage data, but the representation we'll use is TBD.
Singleton values for observed and predicted future usage will rapidly prove inadequate, so we will support the following structure for extended usage information:
Singleton values for observed and predicted future usage will rapidly prove
inadequate, so we will support the following structure for extended usage
information:
```yaml
resourceStatus: [
@ -222,8 +348,12 @@ where a `<CPU-info>` or `<memory-info>` structure looks like this:
}
```
All parts of this structure are optional, although we strongly encourage including quantities for 50, 90, 95, 99, 99.5, and 99.9 percentiles. _[In practice, it will be important to include additional info such as the length of the time window over which the averages are calculated, the confidence level, and information-quality metrics such as the number of dropped or discarded data points.]_
and predicted
All parts of this structure are optional, although we strongly encourage
including quantities for 50, 90, 95, 99, 99.5, and 99.9 percentiles.
_[In practice, it will be important to include additional info such as the
length of the time window over which the averages are calculated, the
confidence level, and information-quality metrics such as the number of dropped
or discarded data points.]_ and predicted
## Future resource types
@ -245,7 +375,10 @@ and predicted
* Units: bytes
* Compressible? no
The amount of secondary storage space available to a container. The main target is local disk drives and SSDs, although this could also be used to qualify remotely-mounted volumes. Specifying whether a resource is a raw disk, an SSD, a disk array, or a file system fronting any of these, is left for future work.
The amount of secondary storage space available to a container. The main target
is local disk drives and SSDs, although this could also be used to qualify
remotely-mounted volumes. Specifying whether a resource is a raw disk, an SSD, a
disk array, or a file system fronting any of these, is left for future work.
### _[future] Storage time_
@ -254,7 +387,9 @@ The amount of secondary storage space available to a container. The main target
* Internal representation: milli-units
* Compressible? yes
This is the amount of time a container spends accessing disk, including actuator and transfer time. A standard disk drive provides 1.0 diskTime seconds per second.
This is the amount of time a container spends accessing disk, including actuator
and transfer time. A standard disk drive provides 1.0 diskTime seconds per
second.
### _[future] Storage operations_

33
docs/design/scheduler_extender.md
View File

@ -34,11 +34,26 @@ Documentation for other releases can be found at
# Scheduler extender
There are three ways to add new scheduling rules (predicates and priority functions) to Kubernetes: (1) by adding these rules to the scheduler and recompiling (described here: https://github.com/kubernetes/kubernetes/blob/master/docs/devel/scheduler.md), (2) implementing your own scheduler process that runs instead of, or alongside of, the standard Kubernetes scheduler, (3) implementing a "scheduler extender" process that the standard Kubernetes scheduler calls out to as a final pass when making scheduling decisions.
There are three ways to add new scheduling rules (predicates and priority
functions) to Kubernetes: (1) by adding these rules to the scheduler and
recompiling (described here:
https://github.com/kubernetes/kubernetes/blob/master/docs/devel/scheduler.md),
(2) implementing your own scheduler process that runs instead of, or alongside
of, the standard Kubernetes scheduler, (3) implementing a "scheduler extender"
process that the standard Kubernetes scheduler calls out to as a final pass when
making scheduling decisions.
This document describes the third approach. This approach is needed for use cases where scheduling decisions need to be made on resources not directly managed by the standard Kubernetes scheduler. The extender helps make scheduling decisions based on such resources. (Note that the three approaches are not mutually exclusive.)
This document describes the third approach. This approach is needed for use
cases where scheduling decisions need to be made on resources not directly
managed by the standard Kubernetes scheduler. The extender helps make scheduling
decisions based on such resources. (Note that the three approaches are not
mutually exclusive.)
When scheduling a pod, the extender allows an external process to filter and prioritize nodes. Two separate http/https calls are issued to the extender, one for "filter" and one for "prioritize" actions. To use the extender, you must create a scheduler policy configuration file. The configuration specifies how to reach the extender, whether to use http or https and the timeout.
When scheduling a pod, the extender allows an external process to filter and
prioritize nodes. Two separate http/https calls are issued to the extender, one
for "filter" and one for "prioritize" actions. To use the extender, you must
create a scheduler policy configuration file. The configuration specifies how to
reach the extender, whether to use http or https and the timeout.
```go
// Holds the parameters used to communicate with the extender. If a verb is unspecified/empty,
@ -94,7 +109,10 @@ A sample scheduler policy file with extender configuration:
}
```
Arguments passed to the FilterVerb endpoint on the extender are the set of nodes filtered through the k8s predicates and the pod. Arguments passed to the PrioritizeVerb endpoint on the extender are the set of nodes filtered through the k8s predicates and extender predicates and the pod.
Arguments passed to the FilterVerb endpoint on the extender are the set of nodes
filtered through the k8s predicates and the pod. Arguments passed to the
PrioritizeVerb endpoint on the extender are the set of nodes filtered through
the k8s predicates and extender predicates and the pod.
```go
// ExtenderArgs represents the arguments needed by the extender to filter/prioritize
@ -107,9 +125,12 @@ type ExtenderArgs struct {
}
```
The "filter" call returns a list of nodes (api.NodeList). The "prioritize" call returns priorities for each node (schedulerapi.HostPriorityList).
The "filter" call returns a list of nodes (api.NodeList). The "prioritize" call
returns priorities for each node (schedulerapi.HostPriorityList).
The "filter" call may prune the set of nodes based on its predicates. Scores returned by the "prioritize" call are added to the k8s scores (computed through its priority functions) and used for final host selection.
The "filter" call may prune the set of nodes based on its predicates. Scores
returned by the "prioritize" call are added to the k8s scores (computed through
its priority functions) and used for final host selection.
Multiple extenders can be configured in the scheduler policy.

397
docs/design/secrets.md
View File

@ -34,15 +34,17 @@ Documentation for other releases can be found at
## Abstract
A proposal for the distribution of [secrets](../user-guide/secrets.md) (passwords, keys, etc) to the Kubelet and to
containers inside Kubernetes using a custom [volume](../user-guide/volumes.md#secrets) type. See the [secrets example](../user-guide/secrets/) for more information.
A proposal for the distribution of [secrets](../user-guide/secrets.md)
(passwords, keys, etc) to the Kubelet and to containers inside Kubernetes using
a custom [volume](../user-guide/volumes.md#secrets) type. See the
[secrets example](../user-guide/secrets/) for more information.
## Motivation
Secrets are needed in containers to access internal resources like the Kubernetes master or
external resources such as git repositories, databases, etc. Users may also want behaviors in the
kubelet that depend on secret data (credentials for image pull from a docker registry) associated
with pods.
Secrets are needed in containers to access internal resources like the
Kubernetes master or external resources such as git repositories, databases,
etc. Users may also want behaviors in the kubelet that depend on secret data
(credentials for image pull from a docker registry) associated with pods.
Goals of this design:
@ -52,114 +54,127 @@ Goals of this design:
## Constraints and Assumptions
* This design does not prescribe a method for storing secrets; storage of secrets should be
pluggable to accommodate different use-cases
* This design does not prescribe a method for storing secrets; storage of
secrets should be pluggable to accommodate different use-cases
* Encryption of secret data and node security are orthogonal concerns
* It is assumed that node and master are secure and that compromising their security could also
compromise secrets:
* If a node is compromised, the only secrets that could potentially be exposed should be the
secrets belonging to containers scheduled onto it
* It is assumed that node and master are secure and that compromising their
security could also compromise secrets:
* If a node is compromised, the only secrets that could potentially be
exposed should be the secrets belonging to containers scheduled onto it
* If the master is compromised, all secrets in the cluster may be exposed
* Secret rotation is an orthogonal concern, but it should be facilitated by this proposal
* A user who can consume a secret in a container can know the value of the secret; secrets must
be provisioned judiciously
* Secret rotation is an orthogonal concern, but it should be facilitated by
this proposal
* A user who can consume a secret in a container can know the value of the
secret; secrets must be provisioned judiciously
## Use Cases
1. As a user, I want to store secret artifacts for my applications and consume them securely in
containers, so that I can keep the configuration for my applications separate from the images
that use them:
1. As a cluster operator, I want to allow a pod to access the Kubernetes master using a custom
`.kubeconfig` file, so that I can securely reach the master
2. As a cluster operator, I want to allow a pod to access a Docker registry using credentials
from a `.dockercfg` file, so that containers can push images
3. As a cluster operator, I want to allow a pod to access a git repository using SSH keys,
so that I can push to and fetch from the repository
2. As a user, I want to allow containers to consume supplemental information about services such
as username and password which should be kept secret, so that I can share secrets about a
service amongst the containers in my application securely
3. As a user, I want to associate a pod with a `ServiceAccount` that consumes a secret and have
the kubelet implement some reserved behaviors based on the types of secrets the service account
consumes:
1. As a user, I want to store secret artifacts for my applications and consume
them securely in containers, so that I can keep the configuration for my
applications separate from the images that use them:
1. As a cluster operator, I want to allow a pod to access the Kubernetes
master using a custom `.kubeconfig` file, so that I can securely reach the
master
2. As a cluster operator, I want to allow a pod to access a Docker registry
using credentials from a `.dockercfg` file, so that containers can push images
3. As a cluster operator, I want to allow a pod to access a git repository
using SSH keys, so that I can push to and fetch from the repository
2. As a user, I want to allow containers to consume supplemental information
about services such as username and password which should be kept secret, so
that I can share secrets about a service amongst the containers in my
application securely
3. As a user, I want to associate a pod with a `ServiceAccount` that consumes a
secret and have the kubelet implement some reserved behaviors based on the types
of secrets the service account consumes:
1. Use credentials for a docker registry to pull the pod's docker image
2. Present Kubernetes auth token to the pod or transparently decorate traffic between the pod
and master service
4. As a user, I want to be able to indicate that a secret expires and for that secret's value to
be rotated once it expires, so that the system can help me follow good practices
2. Present Kubernetes auth token to the pod or transparently decorate
traffic between the pod and master service
4. As a user, I want to be able to indicate that a secret expires and for that
secret's value to be rotated once it expires, so that the system can help me
follow good practices
### Use-Case: Configuration artifacts
Many configuration files contain secrets intermixed with other configuration information. For
example, a user's application may contain a properties file than contains database credentials,
SaaS API tokens, etc. Users should be able to consume configuration artifacts in their containers
and be able to control the path on the container's filesystems where the artifact will be
presented.
Many configuration files contain secrets intermixed with other configuration
information. For example, a user's application may contain a properties file
than contains database credentials, SaaS API tokens, etc. Users should be able
to consume configuration artifacts in their containers and be able to control
the path on the container's filesystems where the artifact will be presented.
### Use-Case: Metadata about services
Most pieces of information about how to use a service are secrets. For example, a service that
provides a MySQL database needs to provide the username, password, and database name to consumers
so that they can authenticate and use the correct database. Containers in pods consuming the MySQL
service would also consume the secrets associated with the MySQL service.
Most pieces of information about how to use a service are secrets. For example,
a service that provides a MySQL database needs to provide the username,
password, and database name to consumers so that they can authenticate and use
the correct database. Containers in pods consuming the MySQL service would also
consume the secrets associated with the MySQL service.
### Use-Case: Secrets associated with service accounts
[Service Accounts](service_accounts.md) are proposed as a
mechanism to decouple capabilities and security contexts from individual human users. A
`ServiceAccount` contains references to some number of secrets. A `Pod` can specify that it is
associated with a `ServiceAccount`. Secrets should have a `Type` field to allow the Kubelet and
other system components to take action based on the secret's type.
[Service Accounts](service_accounts.md) are proposed as a mechanism to decouple
capabilities and security contexts from individual human users. A
`ServiceAccount` contains references to some number of secrets. A `Pod` can
specify that it is associated with a `ServiceAccount`. Secrets should have a
`Type` field to allow the Kubelet and other system components to take action
based on the secret's type.
#### Example: service account consumes auth token secret
As an example, the service account proposal discusses service accounts consuming secrets which
contain Kubernetes auth tokens. When a Kubelet starts a pod associated with a service account
which consumes this type of secret, the Kubelet may take a number of actions:
As an example, the service account proposal discusses service accounts consuming
secrets which contain Kubernetes auth tokens. When a Kubelet starts a pod
associated with a service account which consumes this type of secret, the
Kubelet may take a number of actions:
1. Expose the secret in a `.kubernetes_auth` file in a well-known location in the container's
file system
2. Configure that node's `kube-proxy` to decorate HTTP requests from that pod to the
`kubernetes-master` service with the auth token, e. g. by adding a header to the request
(see the [LOAS Daemon](http://issue.k8s.io/2209) proposal)
1. Expose the secret in a `.kubernetes_auth` file in a well-known location in
the container's file system
2. Configure that node's `kube-proxy` to decorate HTTP requests from that pod
to the `kubernetes-master` service with the auth token, e. g. by adding a header
to the request (see the [LOAS Daemon](http://issue.k8s.io/2209) proposal)
#### Example: service account consumes docker registry credentials
Another example use case is where a pod is associated with a secret containing docker registry
credentials. The Kubelet could use these credentials for the docker pull to retrieve the image.
Another example use case is where a pod is associated with a secret containing
docker registry credentials. The Kubelet could use these credentials for the
docker pull to retrieve the image.
### Use-Case: Secret expiry and rotation
Rotation is considered a good practice for many types of secret data. It should be possible to
express that a secret has an expiry date; this would make it possible to implement a system
component that could regenerate expired secrets. As an example, consider a component that rotates
expired secrets. The rotator could periodically regenerate the values for expired secrets of
common types and update their expiry dates.
Rotation is considered a good practice for many types of secret data. It should
be possible to express that a secret has an expiry date; this would make it
possible to implement a system component that could regenerate expired secrets.
As an example, consider a component that rotates expired secrets. The rotator
could periodically regenerate the values for expired secrets of common types and
update their expiry dates.
## Deferral: Consuming secrets as environment variables
Some images will expect to receive configuration items as environment variables instead of files.
We should consider what the best way to allow this is; there are a few different options:
Some images will expect to receive configuration items as environment variables
instead of files. We should consider what the best way to allow this is; there
are a few different options:
1. Force the user to adapt files into environment variables. Users can store secrets that need to
be presented as environment variables in a format that is easy to consume from a shell:
1. Force the user to adapt files into environment variables. Users can store
secrets that need to be presented as environment variables in a format that is
easy to consume from a shell:
$ cat /etc/secrets/my-secret.txt
export MY_SECRET_ENV=MY_SECRET_VALUE
The user could `source` the file at `/etc/secrets/my-secret` prior to executing the command for
the image either inline in the command or in an init script,
The user could `source` the file at `/etc/secrets/my-secret` prior to
executing the command for the image either inline in the command or in an init
script.
2. Give secrets an attribute that allows users to express the intent that the platform should
generate the above syntax in the file used to present a secret. The user could consume these
files in the same manner as the above option.
2. Give secrets an attribute that allows users to express the intent that the
platform should generate the above syntax in the file used to present a secret.
The user could consume these files in the same manner as the above option.
3. Give secrets attributes that allow the user to express that the secret should be presented to
the container as an environment variable. The container's environment would contain the
desired values and the software in the container could use them without accommodation the
command or setup script.
3. Give secrets attributes that allow the user to express that the secret
should be presented to the container as an environment variable. The container's
environment would contain the desired values and the software in the container
could use them without accommodation the command or setup script.
For our initial work, we will treat all secrets as files to narrow the problem space. There will
be a future proposal that handles exposing secrets as environment variables.
For our initial work, we will treat all secrets as files to narrow the problem
space. There will be a future proposal that handles exposing secrets as
environment variables.
## Flow analysis of secret data with respect to the API server
@ -170,17 +185,19 @@ There are two fundamentally different use-cases for access to secrets:
### Use-Case: CRUD operations by owners
In use cases for CRUD operations, the user experience for secrets should be no different than for
other API resources.
In use cases for CRUD operations, the user experience for secrets should be no
different than for other API resources.
#### Data store backing the REST API
The data store backing the REST API should be pluggable because different cluster operators will
have different preferences for the central store of secret data. Some possibilities for storage:
The data store backing the REST API should be pluggable because different
cluster operators will have different preferences for the central store of
secret data. Some possibilities for storage:
1. An etcd collection alongside the storage for other API resources
2. A collocated [HSM](http://en.wikipedia.org/wiki/Hardware_security_module)
3. A secrets server like [Vault](https://www.vaultproject.io/) or [Keywhiz](https://square.github.io/keywhiz/)
3. A secrets server like [Vault](https://www.vaultproject.io/) or
[Keywhiz](https://square.github.io/keywhiz/)
4. An external datastore such as an external etcd, RDBMS, etc.
#### Size limit for secrets
@ -188,101 +205,116 @@ have different preferences for the central store of secret data. Some possibili
There should be a size limit for secrets in order to:
1. Prevent DOS attacks against the API server
2. Allow kubelet implementations that prevent secret data from touching the node's filesystem
2. Allow kubelet implementations that prevent secret data from touching the
node's filesystem
The size limit should satisfy the following conditions:
1. Large enough to store common artifact types (encryption keypairs, certificates, small
configuration files)
2. Small enough to avoid large impact on node resource consumption (storage, RAM for tmpfs, etc)
1. Large enough to store common artifact types (encryption keypairs,
certificates, small configuration files)
2. Small enough to avoid large impact on node resource consumption (storage,
RAM for tmpfs, etc)
To begin discussion, we propose an initial value for this size limit of **1MB**.
#### Other limitations on secrets
Defining a policy for limitations on how a secret may be referenced by another API resource and how
constraints should be applied throughout the cluster is tricky due to the number of variables
involved:
Defining a policy for limitations on how a secret may be referenced by another
API resource and how constraints should be applied throughout the cluster is
tricky due to the number of variables involved:
1. Should there be a maximum number of secrets a pod can reference via a volume?
1. Should there be a maximum number of secrets a pod can reference via a
volume?
2. Should there be a maximum number of secrets a service account can reference?
3. Should there be a total maximum number of secrets a pod can reference via its own spec and its
associated service account?
4. Should there be a total size limit on the amount of secret data consumed by a pod?
3. Should there be a total maximum number of secrets a pod can reference via
its own spec and its associated service account?
4. Should there be a total size limit on the amount of secret data consumed by
a pod?
5. How will cluster operators want to be able to configure these limits?
6. How will these limits impact API server validations?
7. How will these limits affect scheduling?
For now, we will not implement validations around these limits. Cluster operators will decide how
much node storage is allocated to secrets. It will be the operator's responsibility to ensure that
the allocated storage is sufficient for the workload scheduled onto a node.
For now, we will not implement validations around these limits. Cluster
operators will decide how much node storage is allocated to secrets. It will be
the operator's responsibility to ensure that the allocated storage is sufficient
for the workload scheduled onto a node.
For now, kubelets will only attach secrets to api-sourced pods, and not file- or http-sourced
ones. Doing so would:
For now, kubelets will only attach secrets to api-sourced pods, and not file-
or http-sourced ones. Doing so would:
- confuse the secrets admission controller in the case of mirror pods.
- create an apiserver-liveness dependency -- avoiding this dependency is a main reason to use non-api-source pods.
- create an apiserver-liveness dependency -- avoiding this dependency is a
main reason to use non-api-source pods.
### Use-Case: Kubelet read of secrets for node
The use-case where the kubelet reads secrets has several additional requirements:
1. Kubelets should only be able to receive secret data which is required by pods scheduled onto
the kubelet's node
1. Kubelets should only be able to receive secret data which is required by
pods scheduled onto the kubelet's node
2. Kubelets should have read-only access to secret data
3. Secret data should not be transmitted over the wire insecurely
4. Kubelets must ensure pods do not have access to each other's secrets
#### Read of secret data by the Kubelet
The Kubelet should only be allowed to read secrets which are consumed by pods scheduled onto that
Kubelet's node and their associated service accounts. Authorization of the Kubelet to read this
data would be delegated to an authorization plugin and associated policy rule.
The Kubelet should only be allowed to read secrets which are consumed by pods
scheduled onto that Kubelet's node and their associated service accounts.
Authorization of the Kubelet to read this data would be delegated to an
authorization plugin and associated policy rule.
#### Secret data on the node: data at rest
Consideration must be given to whether secret data should be allowed to be at rest on the node:
Consideration must be given to whether secret data should be allowed to be at
rest on the node:
1. If secret data is not allowed to be at rest, the size of secret data becomes another draw on
the node's RAM - should it affect scheduling?
1. If secret data is not allowed to be at rest, the size of secret data becomes
another draw on the node's RAM - should it affect scheduling?
2. If secret data is allowed to be at rest, should it be encrypted?
1. If so, how should be this be done?
2. If not, what threats exist? What types of secret are appropriate to store this way?
2. If not, what threats exist? What types of secret are appropriate to
store this way?
For the sake of limiting complexity, we propose that initially secret data should not be allowed
to be at rest on a node; secret data should be stored on a node-level tmpfs filesystem. This
filesystem can be subdivided into directories for use by the kubelet and by the volume plugin.
For the sake of limiting complexity, we propose that initially secret data
should not be allowed to be at rest on a node; secret data should be stored on a
node-level tmpfs filesystem. This filesystem can be subdivided into directories
for use by the kubelet and by the volume plugin.
#### Secret data on the node: resource consumption
The Kubelet will be responsible for creating the per-node tmpfs file system for secret storage.
It is hard to make a prescriptive declaration about how much storage is appropriate to reserve for
secrets because different installations will vary widely in available resources, desired pod to
node density, overcommit policy, and other operation dimensions. That being the case, we propose
for simplicity that the amount of secret storage be controlled by a new parameter to the kubelet
with a default value of **64MB**. It is the cluster operator's responsibility to handle choosing
the right storage size for their installation and configuring their Kubelets correctly.
The Kubelet will be responsible for creating the per-node tmpfs file system for
secret storage. It is hard to make a prescriptive declaration about how much
storage is appropriate to reserve for secrets because different installations
will vary widely in available resources, desired pod to node density, overcommit
policy, and other operation dimensions. That being the case, we propose for
simplicity that the amount of secret storage be controlled by a new parameter to
the kubelet with a default value of **64MB**. It is the cluster operator's
responsibility to handle choosing the right storage size for their installation
and configuring their Kubelets correctly.
Configuring each Kubelet is not the ideal story for operator experience; it is more intuitive that
the cluster-wide storage size be readable from a central configuration store like the one proposed
in [#1553](http://issue.k8s.io/1553). When such a store
exists, the Kubelet could be modified to read this configuration item from the store.
Configuring each Kubelet is not the ideal story for operator experience; it is
more intuitive that the cluster-wide storage size be readable from a central
configuration store like the one proposed in [#1553](http://issue.k8s.io/1553).
When such a store exists, the Kubelet could be modified to read this
configuration item from the store.
When the Kubelet is modified to advertise node resources (as proposed in
[#4441](http://issue.k8s.io/4441)), the capacity calculation
for available memory should factor in the potential size of the node-level tmpfs in order to avoid
memory overcommit on the node.
for available memory should factor in the potential size of the node-level tmpfs
in order to avoid memory overcommit on the node.
#### Secret data on the node: isolation
Every pod will have a [security context](security_context.md).
Secret data on the node should be isolated according to the security context of the container. The
Kubelet volume plugin API will be changed so that a volume plugin receives the security context of
a volume along with the volume spec. This will allow volume plugins to implement setting the
security context of volumes they manage.
Secret data on the node should be isolated according to the security context of
the container. The Kubelet volume plugin API will be changed so that a volume
plugin receives the security context of a volume along with the volume spec.
This will allow volume plugins to implement setting the security context of
volumes they manage.
## Community work
Several proposals / upstream patches are notable as background for this proposal:
Several proposals / upstream patches are notable as background for this
proposal:
1. [Docker vault proposal](https://github.com/docker/docker/issues/10310)
2. [Specification for image/container standardization based on volumes](https://github.com/docker/docker/issues/9277)
@ -292,14 +324,15 @@ Several proposals / upstream patches are notable as background for this proposal
## Proposed Design
We propose a new `Secret` resource which is mounted into containers with a new volume type. Secret
volumes will be handled by a volume plugin that does the actual work of fetching the secret and
storing it. Secrets contain multiple pieces of data that are presented as different files within
the secret volume (example: SSH key pair).
We propose a new `Secret` resource which is mounted into containers with a new
volume type. Secret volumes will be handled by a volume plugin that does the
actual work of fetching the secret and storing it. Secrets contain multiple
pieces of data that are presented as different files within the secret volume
(example: SSH key pair).
In order to remove the burden from the end user in specifying every file that a secret consists of,
it should be possible to mount all files provided by a secret with a single `VolumeMount` entry
in the container specification.
In order to remove the burden from the end user in specifying every file that a
secret consists of, it should be possible to mount all files provided by a
secret with a single `VolumeMount` entry in the container specification.
### Secret API Resource
@ -331,27 +364,30 @@ const (
const MaxSecretSize = 1 * 1024 * 1024
```
A Secret can declare a type in order to provide type information to system components that work
with secrets. The default type is `opaque`, which represents arbitrary user-owned data.
A Secret can declare a type in order to provide type information to system
components that work with secrets. The default type is `opaque`, which
represents arbitrary user-owned data.
Secrets are validated against `MaxSecretSize`. The keys in the `Data` field must be valid DNS
subdomains.
Secrets are validated against `MaxSecretSize`. The keys in the `Data` field must
be valid DNS subdomains.
A new REST API and registry interface will be added to accompany the `Secret` resource. The
default implementation of the registry will store `Secret` information in etcd. Future registry
implementations could store the `TypeMeta` and `ObjectMeta` fields in etcd and store the secret
data in another data store entirely, or store the whole object in another data store.
A new REST API and registry interface will be added to accompany the `Secret`
resource. The default implementation of the registry will store `Secret`
information in etcd. Future registry implementations could store the `TypeMeta`
and `ObjectMeta` fields in etcd and store the secret data in another data store
entirely, or store the whole object in another data store.
#### Other validations related to secrets
Initially there will be no validations for the number of secrets a pod references, or the number of
secrets that can be associated with a service account. These may be added in the future as the
finer points of secrets and resource allocation are fleshed out.
Initially there will be no validations for the number of secrets a pod
references, or the number of secrets that can be associated with a service
account. These may be added in the future as the finer points of secrets and
resource allocation are fleshed out.
### Secret Volume Source
A new `SecretSource` type of volume source will be added to the `VolumeSource` struct in the
API:
A new `SecretSource` type of volume source will be added to the `VolumeSource`
struct in the API:
```go
type VolumeSource struct {
@ -366,19 +402,21 @@ type SecretSource struct {
}
```
Secret volume sources are validated to ensure that the specified object reference actually points
to an object of type `Secret`.
Secret volume sources are validated to ensure that the specified object
reference actually points to an object of type `Secret`.
In the future, the `SecretSource` will be extended to allow:
1. Fine-grained control over which pieces of secret data are exposed in the volume
1. Fine-grained control over which pieces of secret data are exposed in the
volume
2. The paths and filenames for how secret data are exposed
### Secret Volume Plugin
A new Kubelet volume plugin will be added to handle volumes with a secret source. This plugin will
require access to the API server to retrieve secret data and therefore the volume `Host` interface
will have to change to expose a client interface:
A new Kubelet volume plugin will be added to handle volumes with a secret
source. This plugin will require access to the API server to retrieve secret
data and therefore the volume `Host` interface will have to change to expose a
client interface:
```go
type Host interface {
@ -394,36 +432,42 @@ The secret volume plugin will be responsible for:
1. Returning a `volume.Mounter` implementation from `NewMounter` that:
1. Retrieves the secret data for the volume from the API server
2. Places the secret data onto the container's filesystem
3. Sets the correct security attributes for the volume based on the pod's `SecurityContext`
2. Returning a `volume.Unmounter` implementation from `NewUnmounter` that cleans the volume from the
container's filesystem
3. Sets the correct security attributes for the volume based on the pod's
`SecurityContext`
2. Returning a `volume.Unmounter` implementation from `NewUnmounter` that
cleans the volume from the container's filesystem
### Kubelet: Node-level secret storage
The Kubelet must be modified to accept a new parameter for the secret storage size and to create
a tmpfs file system of that size to store secret data. Rough accounting of specific changes:
The Kubelet must be modified to accept a new parameter for the secret storage
size and to create a tmpfs file system of that size to store secret data. Rough
accounting of specific changes:
1. The Kubelet should have a new field added called `secretStorageSize`; units are megabytes
1. The Kubelet should have a new field added called `secretStorageSize`; units
are megabytes
2. `NewMainKubelet` should accept a value for secret storage size
3. The Kubelet server should have a new flag added for secret storage size
4. The Kubelet's `setupDataDirs` method should be changed to create the secret storage
4. The Kubelet's `setupDataDirs` method should be changed to create the secret
storage
### Kubelet: New behaviors for secrets associated with service accounts
For use-cases where the Kubelet's behavior is affected by the secrets associated with a pod's
`ServiceAccount`, the Kubelet will need to be changed. For example, if secrets of type
`docker-reg-auth` affect how the pod's images are pulled, the Kubelet will need to be changed
to accommodate this. Subsequent proposals can address this on a type-by-type basis.
For use-cases where the Kubelet's behavior is affected by the secrets associated
with a pod's `ServiceAccount`, the Kubelet will need to be changed. For example,
if secrets of type `docker-reg-auth` affect how the pod's images are pulled, the
Kubelet will need to be changed to accommodate this. Subsequent proposals can
address this on a type-by-type basis.
## Examples
For clarity, let's examine some detailed examples of some common use-cases in terms of the
suggested changes. All of these examples are assumed to be created in a namespace called
`example`.
For clarity, let's examine some detailed examples of some common use-cases in
terms of the suggested changes. All of these examples are assumed to be created
in a namespace called `example`.
### Use-Case: Pod with ssh keys
To create a pod that uses an ssh key stored as a secret, we first need to create a secret:
To create a pod that uses an ssh key stored as a secret, we first need to create
a secret:
```json
{
@ -443,7 +487,8 @@ To create a pod that uses an ssh key stored as a secret, we first need to create
base64 strings. Newlines are not valid within these strings and must be
omitted.
Now we can create a pod which references the secret with the ssh key and consumes it in a volume:
Now we can create a pod which references the secret with the ssh key and
consumes it in a volume:
```json
{
@ -486,7 +531,8 @@ When the container's command runs, the pieces of the key will be available in:
/etc/secret-volume/id-rsa.pub
/etc/secret-volume/id-rsa
The container is then free to use the secret data to establish an ssh connection.
The container is then free to use the secret data to establish an ssh
connection.
### Use-Case: Pods with pod / test credentials
@ -602,8 +648,9 @@ The pods:
}
```
The specs for the two pods differ only in the value of the object referred to by the secret volume
source. Both containers will have the following files present on their filesystems:
The specs for the two pods differ only in the value of the object referred to by
the secret volume source. Both containers will have the following files present
on their filesystems:
/etc/secret-volume/username
/etc/secret-volume/password

194
docs/design/security.md
View File

@ -34,37 +34,57 @@ Documentation for other releases can be found at
# Security in Kubernetes
Kubernetes should define a reasonable set of security best practices that allows processes to be isolated from each other, from the cluster infrastructure, and which preserves important boundaries between those who manage the cluster, and those who use the cluster.
Kubernetes should define a reasonable set of security best practices that allows
processes to be isolated from each other, from the cluster infrastructure, and
which preserves important boundaries between those who manage the cluster, and
those who use the cluster.
While Kubernetes today is not primarily a multi-tenant system, the long term evolution of Kubernetes will increasingly rely on proper boundaries between users and administrators. The code running on the cluster must be appropriately isolated and secured to prevent malicious parties from affecting the entire cluster.
While Kubernetes today is not primarily a multi-tenant system, the long term
evolution of Kubernetes will increasingly rely on proper boundaries between
users and administrators. The code running on the cluster must be appropriately
isolated and secured to prevent malicious parties from affecting the entire
cluster.
## High Level Goals
1. Ensure a clear isolation between the container and the underlying host it runs on
2. Limit the ability of the container to negatively impact the infrastructure or other containers
3. [Principle of Least Privilege](http://en.wikipedia.org/wiki/Principle_of_least_privilege) - ensure components are only authorized to perform the actions they need, and limit the scope of a compromise by limiting the capabilities of individual components
4. Reduce the number of systems that have to be hardened and secured by defining clear boundaries between components
1. Ensure a clear isolation between the container and the underlying host it
runs on
2. Limit the ability of the container to negatively impact the infrastructure
or other containers
3. [Principle of Least Privilege](http://en.wikipedia.org/wiki/Principle_of_least_privilege) -
ensure components are only authorized to perform the actions they need, and
limit the scope of a compromise by limiting the capabilities of individual
components
4. Reduce the number of systems that have to be hardened and secured by
defining clear boundaries between components
5. Allow users of the system to be cleanly separated from administrators
6. Allow administrative functions to be delegated to users where necessary
7. Allow applications to be run on the cluster that have "secret" data (keys, certs, passwords) which is properly abstracted from "public" data.
7. Allow applications to be run on the cluster that have "secret" data (keys,
certs, passwords) which is properly abstracted from "public" data.
## Use cases
### Roles
We define "user" as a unique identity accessing the Kubernetes API server, which may be a human or an automated process. Human users fall into the following categories:
We define "user" as a unique identity accessing the Kubernetes API server, which
may be a human or an automated process. Human users fall into the following
categories:
1. k8s admin - administers a Kubernetes cluster and has access to the underlying components of the system
2. k8s project administrator - administrates the security of a small subset of the cluster
3. k8s developer - launches pods on a Kubernetes cluster and consumes cluster resources
1. k8s admin - administers a Kubernetes cluster and has access to the underlying
components of the system
2. k8s project administrator - administrates the security of a small subset of
the cluster
3. k8s developer - launches pods on a Kubernetes cluster and consumes cluster
resources
Automated process users fall into the following categories:
1. k8s container user - a user that processes running inside a container (on the cluster) can use to access other cluster resources independent of the human users attached to a project
2. k8s infrastructure user - the user that Kubernetes infrastructure components use to perform cluster functions with clearly defined roles
1. k8s container user - a user that processes running inside a container (on the
cluster) can use to access other cluster resources independent of the human
users attached to a project
2. k8s infrastructure user - the user that Kubernetes infrastructure components
use to perform cluster functions with clearly defined roles
### Description of roles
@ -73,9 +93,11 @@ Automated process users fall into the following categories:
* making some of their own images, and using some "community" docker images
* know which pods need to talk to which other pods
* decide which pods should share files with other pods, and which should not.
* reason about application level security, such as containing the effects of a local-file-read exploit in a webserver pod.
* reason about application level security, such as containing the effects of a
local-file-read exploit in a webserver pod.
* do not often reason about operating system or organizational security.
* are not necessarily comfortable reasoning about the security properties of a system at the level of detail of Linux Capabilities, SELinux, AppArmor, etc.
* are not necessarily comfortable reasoning about the security properties of a
system at the level of detail of Linux Capabilities, SELinux, AppArmor, etc.
* Project Admins:
* allocate identity and roles within a namespace
@ -85,44 +107,81 @@ Automated process users fall into the following categories:
* are less focused about application security
* Administrators:
* are less focused on application security. Focused on operating system security.
* protect the node from bad actors in containers, and properly-configured innocent containers from bad actors in other containers.
* comfortable reasoning about the security properties of a system at the level of detail of Linux Capabilities, SELinux, AppArmor, etc.
* decides who can use which Linux Capabilities, run privileged containers, use hostPath, etc.
* e.g. a team that manages Ceph or a mysql server might be trusted to have raw access to storage devices in some organizations, but teams that develop the applications at higher layers would not.
* are less focused on application security. Focused on operating system
security.
* protect the node from bad actors in containers, and properly-configured
innocent containers from bad actors in other containers.
* comfortable reasoning about the security properties of a system at the level
of detail of Linux Capabilities, SELinux, AppArmor, etc.
* decides who can use which Linux Capabilities, run privileged containers, use
hostPath, etc.
* e.g. a team that manages Ceph or a mysql server might be trusted to have
raw access to storage devices in some organizations, but teams that develop the
applications at higher layers would not.
## Proposed Design
A pod runs in a *security context* under a *service account* that is defined by an administrator or project administrator, and the *secrets* a pod has access to is limited by that *service account*.
A pod runs in a *security context* under a *service account* that is defined by
an administrator or project administrator, and the *secrets* a pod has access to
is limited by that *service account*.
1. The API should authenticate and authorize user actions [authn and authz](access.md)
2. All infrastructure components (kubelets, kube-proxies, controllers, scheduler) should have an infrastructure user that they can authenticate with and be authorized to perform only the functions they require against the API.
3. Most infrastructure components should use the API as a way of exchanging data and changing the system, and only the API should have access to the underlying data store (etcd)
4. When containers run on the cluster and need to talk to other containers or the API server, they should be identified and authorized clearly as an autonomous process via a [service account](service_accounts.md)
1. If the user who started a long-lived process is removed from access to the cluster, the process should be able to continue without interruption
2. If the user who started processes are removed from the cluster, administrators may wish to terminate their processes in bulk
3. When containers run with a service account, the user that created / triggered the service account behavior must be associated with the container's action
5. When container processes run on the cluster, they should run in a [security context](security_context.md) that isolates those processes via Linux user security, user namespaces, and permissions.
1. Administrators should be able to configure the cluster to automatically confine all container processes as a non-root, randomly assigned UID
2. Administrators should be able to ensure that container processes within the same namespace are all assigned the same unix user UID
3. Administrators should be able to limit which developers and project administrators have access to higher privilege actions
4. Project administrators should be able to run pods within a namespace under different security contexts, and developers must be able to specify which of the available security contexts they may use
5. Developers should be able to run their own images or images from the community and expect those images to run correctly
6. Developers may need to ensure their images work within higher security requirements specified by administrators
7. When available, Linux kernel user namespaces can be used to ensure 5.2 and 5.4 are met.
8. When application developers want to share filesystem data via distributed filesystems, the Unix user ids on those filesystems must be consistent across different container processes
6. Developers should be able to define [secrets](secrets.md) that are automatically added to the containers when pods are run
1. Secrets are files injected into the container whose values should not be displayed within a pod. Examples:
2. All infrastructure components (kubelets, kube-proxies, controllers,
scheduler) should have an infrastructure user that they can authenticate with
and be authorized to perform only the functions they require against the API.
3. Most infrastructure components should use the API as a way of exchanging data
and changing the system, and only the API should have access to the underlying
data store (etcd)
4. When containers run on the cluster and need to talk to other containers or
the API server, they should be identified and authorized clearly as an
autonomous process via a [service account](service_accounts.md)
1. If the user who started a long-lived process is removed from access to
the cluster, the process should be able to continue without interruption
2. If the user who started processes are removed from the cluster,
administrators may wish to terminate their processes in bulk
3. When containers run with a service account, the user that created /
triggered the service account behavior must be associated with the container's
action
5. When container processes run on the cluster, they should run in a
[security context](security_context.md) that isolates those processes via Linux
user security, user namespaces, and permissions.
1. Administrators should be able to configure the cluster to automatically
confine all container processes as a non-root, randomly assigned UID
2. Administrators should be able to ensure that container processes within
the same namespace are all assigned the same unix user UID
3. Administrators should be able to limit which developers and project
administrators have access to higher privilege actions
4. Project administrators should be able to run pods within a namespace
under different security contexts, and developers must be able to specify which
of the available security contexts they may use
5. Developers should be able to run their own images or images from the
community and expect those images to run correctly
6. Developers may need to ensure their images work within higher security
requirements specified by administrators
7. When available, Linux kernel user namespaces can be used to ensure 5.2
and 5.4 are met.
8. When application developers want to share filesystem data via distributed
filesystems, the Unix user ids on those filesystems must be consistent across
different container processes
6. Developers should be able to define [secrets](secrets.md) that are
automatically added to the containers when pods are run
1. Secrets are files injected into the container whose values should not be
displayed within a pod. Examples:
1. An SSH private key for git cloning remote data
2. A client certificate for accessing a remote system
3. A private key and certificate for a web server
4. A .kubeconfig file with embedded cert / token data for accessing the Kubernetes master
4. A .kubeconfig file with embedded cert / token data for accessing the
Kubernetes master
5. A .dockercfg file for pulling images from a protected registry
2. Developers should be able to define the pod spec so that a secret lands in a specific location
3. Project administrators should be able to limit developers within a namespace from viewing or modifying secrets (anyone who can launch an arbitrary pod can view secrets)
4. Secrets are generally not copied from one namespace to another when a developer's application definitions are copied
2. Developers should be able to define the pod spec so that a secret lands
in a specific location
3. Project administrators should be able to limit developers within a
namespace from viewing or modifying secrets (anyone who can launch an arbitrary
pod can view secrets)
4. Secrets are generally not copied from one namespace to another when a
developer's application definitions are copied
### Related design discussion
@ -140,15 +199,52 @@ A pod runs in a *security context* under a *service account* that is defined by
### Isolate the data store from the nodes and supporting infrastructure
Access to the central data store (etcd) in Kubernetes allows an attacker to run arbitrary containers on hosts, to gain access to any protected information stored in either volumes or in pods (such as access tokens or shared secrets provided as environment variables), to intercept and redirect traffic from running services by inserting middlemen, or to simply delete the entire history of the custer.
Access to the central data store (etcd) in Kubernetes allows an attacker to run
arbitrary containers on hosts, to gain access to any protected information
stored in either volumes or in pods (such as access tokens or shared secrets
provided as environment variables), to intercept and redirect traffic from
running services by inserting middlemen, or to simply delete the entire history
of the custer.
As a general principle, access to the central data store should be restricted to the components that need full control over the system and which can apply appropriate authorization and authentication of change requests. In the future, etcd may offer granular access control, but that granularity will require an administrator to understand the schema of the data to properly apply security. An administrator must be able to properly secure Kubernetes at a policy level, rather than at an implementation level, and schema changes over time should not risk unintended security leaks.
As a general principle, access to the central data store should be restricted to
the components that need full control over the system and which can apply
appropriate authorization and authentication of change requests. In the future,
etcd may offer granular access control, but that granularity will require an
administrator to understand the schema of the data to properly apply security.
An administrator must be able to properly secure Kubernetes at a policy level,
rather than at an implementation level, and schema changes over time should not
risk unintended security leaks.
Both the Kubelet and Kube Proxy need information related to their specific roles - for the Kubelet, the set of pods it should be running, and for the Proxy, the set of services and endpoints to load balance. The Kubelet also needs to provide information about running pods and historical termination data. The access pattern for both Kubelet and Proxy to load their configuration is an efficient "wait for changes" request over HTTP. It should be possible to limit the Kubelet and Proxy to only access the information they need to perform their roles and no more.
Both the Kubelet and Kube Proxy need information related to their specific roles -
for the Kubelet, the set of pods it should be running, and for the Proxy, the
set of services and endpoints to load balance. The Kubelet also needs to provide
information about running pods and historical termination data. The access
pattern for both Kubelet and Proxy to load their configuration is an efficient
"wait for changes" request over HTTP. It should be possible to limit the Kubelet
and Proxy to only access the information they need to perform their roles and no
more.
The controller manager for Replication Controllers and other future controllers act on behalf of a user via delegation to perform automated maintenance on Kubernetes resources. Their ability to access or modify resource state should be strictly limited to their intended duties and they should be prevented from accessing information not pertinent to their role. For example, a replication controller needs only to create a copy of a known pod configuration, to determine the running state of an existing pod, or to delete an existing pod that it created - it does not need to know the contents or current state of a pod, nor have access to any data in the pods attached volumes.
The controller manager for Replication Controllers and other future controllers
act on behalf of a user via delegation to perform automated maintenance on
Kubernetes resources. Their ability to access or modify resource state should be
strictly limited to their intended duties and they should be prevented from
accessing information not pertinent to their role. For example, a replication
controller needs only to create a copy of a known pod configuration, to
determine the running state of an existing pod, or to delete an existing pod
that it created - it does not need to know the contents or current state of a
pod, nor have access to any data in the pods attached volumes.
The Kubernetes pod scheduler is responsible for reading data from the pod to fit it onto a node in the cluster. At a minimum, it needs access to view the ID of a pod (to craft the binding), its current state, any resource information necessary to identify placement, and other data relevant to concerns like anti-affinity, zone or region preference, or custom logic. It does not need the ability to modify pods or see other resources, only to create bindings. It should not need the ability to delete bindings unless the scheduler takes control of relocating components on failed hosts (which could be implemented by a separate component that can delete bindings but not create them). The scheduler may need read access to user or project-container information to determine preferential location (underspecified at this time).
The Kubernetes pod scheduler is responsible for reading data from the pod to fit
it onto a node in the cluster. At a minimum, it needs access to view the ID of a
pod (to craft the binding), its current state, any resource information
necessary to identify placement, and other data relevant to concerns like
anti-affinity, zone or region preference, or custom logic. It does not need the
ability to modify pods or see other resources, only to create bindings. It
should not need the ability to delete bindings unless the scheduler takes
control of relocating components on failed hosts (which could be implemented by
a separate component that can delete bindings but not create them). The
scheduler may need read access to user or project-container information to
determine preferential location (underspecified at this time).
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

109
docs/design/security_context.md
View File

@ -36,41 +36,59 @@ Documentation for other releases can be found at
## Abstract
A security context is a set of constraints that are applied to a container in order to achieve the following goals (from [security design](security.md)):
A security context is a set of constraints that are applied to a container in
order to achieve the following goals (from [security design](security.md)):
1. Ensure a clear isolation between container and the underlying host it runs on
2. Limit the ability of the container to negatively impact the infrastructure or other containers
1. Ensure a clear isolation between container and the underlying host it runs
on
2. Limit the ability of the container to negatively impact the infrastructure
or other containers
## Background
The problem of securing containers in Kubernetes has come up [before](http://issue.k8s.io/398) and the potential problems with container security are [well known](http://opensource.com/business/14/7/docker-security-selinux). Although it is not possible to completely isolate Docker containers from their hosts, new features like [user namespaces](https://github.com/docker/libcontainer/pull/304) make it possible to greatly reduce the attack surface.
The problem of securing containers in Kubernetes has come up
[before](http://issue.k8s.io/398) and the potential problems with container
security are [well known](http://opensource.com/business/14/7/docker-security-selinux).
Although it is not possible to completely isolate Docker containers from their
hosts, new features like [user namespaces](https://github.com/docker/libcontainer/pull/304)
make it possible to greatly reduce the attack surface.
## Motivation
### Container isolation
In order to improve container isolation from host and other containers running on the host, containers should only be
granted the access they need to perform their work. To this end it should be possible to take advantage of Docker
features such as the ability to [add or remove capabilities](https://docs.docker.com/reference/run/#runtime-privilege-linux-capabilities-and-lxc-configuration) and [assign MCS labels](https://docs.docker.com/reference/run/#security-configuration)
In order to improve container isolation from host and other containers running
on the host, containers should only be granted the access they need to perform
their work. To this end it should be possible to take advantage of Docker
features such as the ability to
[add or remove capabilities](https://docs.docker.com/reference/run/#runtime-privilege-linux-capabilities-and-lxc-configuration)
and [assign MCS labels](https://docs.docker.com/reference/run/#security-configuration)
to the container process.
Support for user namespaces has recently been [merged](https://github.com/docker/libcontainer/pull/304) into Docker's libcontainer project and should soon surface in Docker itself. It will make it possible to assign a range of unprivileged uids and gids from the host to each container, improving the isolation between host and container and between containers.
Support for user namespaces has recently been
[merged](https://github.com/docker/libcontainer/pull/304) into Docker's
libcontainer project and should soon surface in Docker itself. It will make it
possible to assign a range of unprivileged uids and gids from the host to each
container, improving the isolation between host and container and between
containers.
### External integration with shared storage
In order to support external integration with shared storage, processes running in a Kubernetes cluster
should be able to be uniquely identified by their Unix UID, such that a chain of ownership can be established.
Processes in pods will need to have consistent UID/GID/SELinux category labels in order to access shared disks.
In order to support external integration with shared storage, processes running
in a Kubernetes cluster should be able to be uniquely identified by their Unix
UID, such that a chain of ownership can be established. Processes in pods will
need to have consistent UID/GID/SELinux category labels in order to access
shared disks.
## Constraints and Assumptions
* It is out of the scope of this document to prescribe a specific set
of constraints to isolate containers from their host. Different use cases need different
settings.
* The concept of a security context should not be tied to a particular security mechanism or platform
(ie. SELinux, AppArmor)
* Applying a different security context to a scope (namespace or pod) requires a solution such as the one proposed for
[service accounts](service_accounts.md).
* It is out of the scope of this document to prescribe a specific set of
constraints to isolate containers from their host. Different use cases need
different settings.
* The concept of a security context should not be tied to a particular security
mechanism or platform (ie. SELinux, AppArmor)
* Applying a different security context to a scope (namespace or pod) requires
a solution such as the one proposed for [service accounts](service_accounts.md).
## Use Cases
@ -78,47 +96,51 @@ In order of increasing complexity, following are example use cases that would
be addressed with security contexts:
1. Kubernetes is used to run a single cloud application. In order to protect
nodes from containers:
nodes from containers:
* All containers run as a single non-root user
* Privileged containers are disabled
* All containers run with a particular MCS label
* Kernel capabilities like CHOWN and MKNOD are removed from containers
2. Just like case #1, except that I have more than one application running on
the Kubernetes cluster.
the Kubernetes cluster.
* Each application is run in its own namespace to avoid name collisions
* For each application a different uid and MCS label is used
3. Kubernetes is used as the base for a PAAS with
multiple projects, each project represented by a namespace.
3. Kubernetes is used as the base for a PAAS with multiple projects, each
project represented by a namespace.
* Each namespace is associated with a range of uids/gids on the node that
are mapped to uids/gids on containers using linux user namespaces.
are mapped to uids/gids on containers using linux user namespaces.
* Certain pods in each namespace have special privileges to perform system
actions such as talking back to the server for deployment, run docker
builds, etc.
actions such as talking back to the server for deployment, run docker builds,
etc.
* External NFS storage is assigned to each namespace and permissions set
using the range of uids/gids assigned to that namespace.
using the range of uids/gids assigned to that namespace.
## Proposed Design
### Overview
A *security context* consists of a set of constraints that determine how a container
is secured before getting created and run. A security context resides on the container and represents the runtime parameters that will
be used to create and run the container via container APIs. A *security context provider* is passed to the Kubelet so it can have a chance
to mutate Docker API calls in order to apply the security context.
A *security context* consists of a set of constraints that determine how a
container is secured before getting created and run. A security context resides
on the container and represents the runtime parameters that will be used to
create and run the container via container APIs. A *security context provider*
is passed to the Kubelet so it can have a chance to mutate Docker API calls in
order to apply the security context.
It is recommended that this design be implemented in two phases:
1. Implement the security context provider extension point in the Kubelet
so that a default security context can be applied on container run and creation.
so that a default security context can be applied on container run and creation.
2. Implement a security context structure that is part of a service account. The
default context provider can then be used to apply a security context based
on the service account associated with the pod.
default context provider can then be used to apply a security context based on
the service account associated with the pod.
### Security Context Provider
The Kubelet will have an interface that points to a `SecurityContextProvider`. The `SecurityContextProvider` is invoked before creating and running a given container:
The Kubelet will have an interface that points to a `SecurityContextProvider`.
The `SecurityContextProvider` is invoked before creating and running a given
container:
```go
type SecurityContextProvider interface {
@ -138,12 +160,14 @@ type SecurityContextProvider interface {
}
```
If the value of the SecurityContextProvider field on the Kubelet is nil, the kubelet will create and run the container as it does today.
If the value of the SecurityContextProvider field on the Kubelet is nil, the
kubelet will create and run the container as it does today.
### Security Context
A security context resides on the container and represents the runtime parameters that will
be used to create and run the container via container APIs. Following is an example of an initial implementation:
A security context resides on the container and represents the runtime
parameters that will be used to create and run the container via container APIs.
Following is an example of an initial implementation:
```go
type Container struct {
@ -189,11 +213,12 @@ type SELinuxOptions struct {
### Admission
It is up to an admission plugin to determine if the security context is acceptable or not. At the
time of writing, the admission control plugin for security contexts will only allow a context that
has defined capabilities or privileged. Contexts that attempt to define a UID or SELinux options
will be denied by default. In the future the admission plugin will base this decision upon
configurable policies that reside within the [service account](http://pr.k8s.io/2297).
It is up to an admission plugin to determine if the security context is
acceptable or not. At the time of writing, the admission control plugin for
security contexts will only allow a context that has defined capabilities or
privileged. Contexts that attempt to define a UID or SELinux options will be
denied by default. In the future the admission plugin will base this decision
upon configurable policies that reside within the [service account](http://pr.k8s.io/2297).
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

147
docs/design/selector-generation.md
View File

@ -37,40 +37,64 @@ Design
# Goals
Make it really hard to accidentally create a job which has an overlapping selector, while still making it possible to chose an arbitrary selector, and without adding complex constraint solving to the APIserver.
Make it really hard to accidentally create a job which has an overlapping
selector, while still making it possible to chose an arbitrary selector, and
without adding complex constraint solving to the APIserver.
# Use Cases
1. user can leave all label and selector fields blank and system will fill in reasonable ones: non-overlappingness guaranteed.
2. user can put on the pod template some labels that are useful to the user, without reasoning about non-overlappingness. System adds additional label to assure not overlapping.
3. If user wants to reparent pods to new job (very rare case) and knows what they are doing, they can completely disable this behavior and specify explicit selector.
4. If a controller that makes jobs, like scheduled job, wants to use different labels, such as the time and date of the run, it can do that.
5. If User reads v1beta1 documentation or reuses v1beta1 Job definitions and just changes the API group, the user should not automatically be allowed to specify a selector, since this is very rarely what people want to do and is error prone.
6. If User downloads an existing job definition, e.g. with `kubectl get jobs/old -o yaml` and tries to modify and post it, he should not create an overlapping job.
7. If User downloads an existing job definition, e.g. with `kubectl get jobs/old -o yaml` and tries to modify and post it, and he accidentally copies the uniquifying label from the old one, then he should not get an error from a label-key conflict, nor get erratic behavior.
8. If user reads swagger docs and sees the selector field, he should not be able to set it without realizing the risks.
8. (Deferred requirement:) If user wants to specify a preferred name for the non-overlappingness key, they can pick a name.
1. user can leave all label and selector fields blank and system will fill in
reasonable ones: non-overlappingness guaranteed.
2. user can put on the pod template some labels that are useful to the user,
without reasoning about non-overlappingness. System adds additional label to
assure not overlapping.
3. If user wants to reparent pods to new job (very rare case) and knows what
they are doing, they can completely disable this behavior and specify explicit
selector.
4. If a controller that makes jobs, like scheduled job, wants to use different
labels, such as the time and date of the run, it can do that.
5. If User reads v1beta1 documentation or reuses v1beta1 Job definitions and
just changes the API group, the user should not automatically be allowed to
specify a selector, since this is very rarely what people want to do and is
error prone.
6. If User downloads an existing job definition, e.g. with
`kubectl get jobs/old -o yaml` and tries to modify and post it, he should not
create an overlapping job.
7. If User downloads an existing job definition, e.g. with
`kubectl get jobs/old -o yaml` and tries to modify and post it, and he
accidentally copies the uniquifying label from the old one, then he should not
get an error from a label-key conflict, nor get erratic behavior.
8. If user reads swagger docs and sees the selector field, he should not be able
to set it without realizing the risks.
8. (Deferred requirement:) If user wants to specify a preferred name for the
non-overlappingness key, they can pick a name.
# Proposed changes
## API
`extensions/v1beta1 Job` remains the same. `batch/v1 Job` changes change as follows.
`extensions/v1beta1 Job` remains the same. `batch/v1 Job` changes change as
follows.
Field `job.spec.manualSelector` is added. It controls whether selectors are automatically
generated. In automatic mode, user cannot make the mistake of creating non-unique selectors.
In manual mode, certain rare use cases are supported.
Field `job.spec.manualSelector` is added. It controls whether selectors are
automatically generated. In automatic mode, user cannot make the mistake of
creating non-unique selectors. In manual mode, certain rare use cases are
supported.
Validation is not changed. A selector must be provided, and it must select the pod template.
Validation is not changed. A selector must be provided, and it must select the
pod template.
Defaulting changes. Defaulting happens in one of two modes:
Defaulting changes. Defaulting happens in one of two modes:
### Automatic Mode
- User does not specify `job.spec.selector`.
- User is probably unaware of the `job.spec.manualSelector` field and does not think about it.
- User optionally puts labels on pod template (optional). user does not think about uniqueness, just labeling for user's own reasons.
- Defaulting logic sets `job.spec.selector` to `matchLabels["controller-uid"]="$UIDOFJOB"`
- User is probably unaware of the `job.spec.manualSelector` field and does not
think about it.
- User optionally puts labels on pod template (optional). User does not think
about uniqueness, just labeling for user's own reasons.
- Defaulting logic sets `job.spec.selector` to
`matchLabels["controller-uid"]="$UIDOFJOB"`
- Defaulting logic appends 2 labels to the `.spec.template.metadata.labels`.
- The first label is controller-uid=$UIDOFJOB.
- The second label is "job-name=$NAMEOFJOB".
@ -80,19 +104,30 @@ Defaulting changes. Defaulting happens in one of two modes:
- User means User or Controller for the rest of this list.
- User does specify `job.spec.selector`.
- User does specify `job.spec.manualSelector=true`
- User puts a unique label or label(s) on pod template (required). user does think carefully about uniqueness.
- User puts a unique label or label(s) on pod template (required). User does
think carefully about uniqueness.
- No defaulting of pod labels or the selector happen.
### Rationale
UID is better than Name in that:
- it allows cross-namespace control someday if we need it.
- it is unique across all kinds. `controller-name=foo` does not ensure uniqueness across Kinds `job` vs `replicaSet`. Even `job-name=foo` has a problem: you might have a `batch.Job` and a `snazzyjob.io/types.Job` -- the latter cannot use label `job-name=foo`, though there is a temptation to do so.
- it uniquely identifies the controller across time. This prevents the case where, for example, someone deletes a job via the REST api or client (where cascade=false), leaving pods around. We don't want those to be picked up unintentionally. It also prevents the case where a user looks at an old job that finished but is not deleted, and tries to select its pods, and gets the wrong impression that it is still running.
- it is unique across all kinds. `controller-name=foo` does not ensure
uniqueness across Kinds `job` vs `replicaSet`. Even `job-name=foo` has a
problem: you might have a `batch.Job` and a `snazzyjob.io/types.Job` -- the
latter cannot use label `job-name=foo`, though there is a temptation to do so.
- it uniquely identifies the controller across time. This prevents the case
where, for example, someone deletes a job via the REST api or client
(where cascade=false), leaving pods around. We don't want those to be picked up
unintentionally. It also prevents the case where a user looks at an old job that
finished but is not deleted, and tries to select its pods, and gets the wrong
impression that it is still running.
Job name is more user friendly. It is self documenting
Commands like `kubectl get pods -l job-name=myjob` should do exactly what is wanted 99.9% of the time. Automated control loops should still use the controller-uid=label.
Commands like `kubectl get pods -l job-name=myjob` should do exactly what is
wanted 99.9% of the time. Automated control loops should still use the
controller-uid=label.
Using both gets the benefits of both, at the cost of some label verbosity.
@ -102,11 +137,15 @@ users looking at a stored pod spec do not need to be aware of this field.
### Overriding Unique Labels
If user does specify `job.spec.selector` then the user must also specify `job.spec.manualSelector`.
This ensures the user knows that what he is doing is not the normal thing to do.
If user does specify `job.spec.selector` then the user must also specify
`job.spec.manualSelector`. This ensures the user knows that what he is doing is
not the normal thing to do.
To prevent users from copying the `job.spec.manualSelector` flag from existing jobs, it will be
optional and default to false, which means when you ask GET and existing job back that didn't use this feature, you don't even see the `job.spec.manualSelector` flag, so you are not tempted to wonder if you should fiddle with it.
To prevent users from copying the `job.spec.manualSelector` flag from existing
jobs, it will be optional and default to false, which means when you ask GET and
existing job back that didn't use this feature, you don't even see the
`job.spec.manualSelector` flag, so you are not tempted to wonder if you should
fiddle with it.
## Job Controller
@ -114,8 +153,8 @@ No changes
## Kubectl
No required changes.
Suggest moving SELECTOR to wide output of `kubectl get jobs` since users do not write the selector.
No required changes. Suggest moving SELECTOR to wide output of `kubectl get
jobs` since users do not write the selector.
## Docs
@ -124,42 +163,50 @@ Recommend `kubectl get jobs -l job-name=name` as the way to find pods of a job.
# Conversion
The following applies to Job, as well as to other types that adopt this pattern.
The following applies to Job, as well as to other types that adopt this pattern:
- Type `extensions/v1beta1` gets a field called `job.spec.autoSelector`.
- Both the internal type and the `batch/v1` type will get `job.spec.manualSelector`.
- Both the internal type and the `batch/v1` type will get
`job.spec.manualSelector`.
- The fields `manualSelector` and `autoSelector` have opposite meanings.
- Each field defaults to false when unset, and so v1beta1 has a different default than v1 and internal. This is intentional: we want new
uses to default to the less error-prone behavior, and we do not want to change the behavior
of v1beta1.
- Each field defaults to false when unset, and so v1beta1 has a different
default than v1 and internal. This is intentional: we want new uses to default
to the less error-prone behavior, and we do not want to change the behavior of
v1beta1.
*Note*: since the internal default is changing, client
library consumers that create Jobs may need to add "job.spec.manualSelector=true" to keep working, or switch
to auto selectors.
*Note*: since the internal default is changing, client library consumers that
create Jobs may need to add "job.spec.manualSelector=true" to keep working, or
switch to auto selectors.
Conversion is as follows:
- `extensions/__internal` to `extensions/v1beta1`: the value of `__internal.Spec.ManualSelector` is defaulted to false if nil, negated, defaulted to nil if false, and written `v1beta1.Spec.AutoSelector`.
- `extensions/v1beta1` to `extensions/__internal`: the value of `v1beta1.SpecAutoSelector` is defaulted to false if nil, negated, defaulted to nil if false, and written to `__internal.Spec.ManualSelector`.
- `extensions/__internal` to `extensions/v1beta1`: the value of
`__internal.Spec.ManualSelector` is defaulted to false if nil, negated,
defaulted to nil if false, and written `v1beta1.Spec.AutoSelector`.
- `extensions/v1beta1` to `extensions/__internal`: the value of
`v1beta1.SpecAutoSelector` is defaulted to false if nil, negated, defaulted to
nil if false, and written to `__internal.Spec.ManualSelector`.
This conversion gives the following properties.
1. Users that previously used v1beta1 do not start seeing a new field when they get back objects.
2. Distinction between originally unset versus explicitly set to false is not preserved (would have been nice to do so, but requires more complicated
solution).
3. Users who only created v1beta1 examples or v1 examples, will not ever see the existence of either field.
4. Since v1beta1 are convertable to/from v1, the storage location (path in etcd) does not need to change, allowing scriptable rollforward/rollback.
1. Users that previously used v1beta1 do not start seeing a new field when they
get back objects.
2. Distinction between originally unset versus explicitly set to false is not
preserved (would have been nice to do so, but requires more complicated
solution).
3. Users who only created v1beta1 examples or v1 examples, will not ever see the
existence of either field.
4. Since v1beta1 are convertable to/from v1, the storage location (path in etcd)
does not need to change, allowing scriptable rollforward/rollback.
# Future Work
Follow this pattern for Deployments, ReplicaSet, DaemonSet when going to v1, if it works well for job.
Follow this pattern for Deployments, ReplicaSet, DaemonSet when going to v1, if
it works well for job.
Docs will be edited to show examples without a `job.spec.selector`.
We probably want as much as possible the same behavior for Job and ReplicationController.
We probably want as much as possible the same behavior for Job and
ReplicationController.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

200
docs/design/service_accounts.md
View File

@ -40,26 +40,30 @@ Processes in Pods may need to call the Kubernetes API. For example:
- scheduler
- replication controller
- node controller
- a map-reduce type framework which has a controller that then tries to make a dynamically determined number of workers and watch them
- a map-reduce type framework which has a controller that then tries to make a
dynamically determined number of workers and watch them
- continuous build and push system
- monitoring system
They also may interact with services other than the Kubernetes API, such as:
- an image repository, such as docker -- both when the images are pulled to start the containers, and for writing
images in the case of pods that generate images.
- accessing other cloud services, such as blob storage, in the context of a large, integrated, cloud offering (hosted
or private).
- an image repository, such as docker -- both when the images are pulled to
start the containers, and for writing images in the case of pods that generate
images.
- accessing other cloud services, such as blob storage, in the context of a
large, integrated, cloud offering (hosted or private).
- accessing files in an NFS volume attached to the pod
## Design Overview
A service account binds together several things:
- a *name*, understood by users, and perhaps by peripheral systems, for an identity
- a *name*, understood by users, and perhaps by peripheral systems, for an
identity
- a *principal* that can be authenticated and [authorized](../admin/authorization.md)
- a [security context](security_context.md), which defines the Linux Capabilities, User IDs, Groups IDs, and other
capabilities and controls on interaction with the file system and OS.
- a set of [secrets](secrets.md), which a container may use to
access various networked resources.
- a [security context](security_context.md), which defines the Linux
Capabilities, User IDs, Groups IDs, and other capabilities and controls on
interaction with the file system and OS.
- a set of [secrets](secrets.md), which a container may use to access various
networked resources.
## Design Discussion
@ -76,94 +80,119 @@ type ServiceAccount struct {
}
```
The name ServiceAccount is chosen because it is widely used already (e.g. by Kerberos and LDAP)
to refer to this type of account. Note that it has no relation to Kubernetes Service objects.
The name ServiceAccount is chosen because it is widely used already (e.g. by
Kerberos and LDAP) to refer to this type of account. Note that it has no
relation to Kubernetes Service objects.
The ServiceAccount object does not include any information that could not be defined separately:
The ServiceAccount object does not include any information that could not be
defined separately:
- username can be defined however users are defined.
- securityContext and secrets are only referenced and are created using the REST API.
- securityContext and secrets are only referenced and are created using the
REST API.
The purpose of the serviceAccount object is twofold:
- to bind usernames to securityContexts and secrets, so that the username can be used to refer succinctly
in contexts where explicitly naming securityContexts and secrets would be inconvenient
- to provide an interface to simplify allocation of new securityContexts and secrets.
- to bind usernames to securityContexts and secrets, so that the username can
be used to refer succinctly in contexts where explicitly naming securityContexts
and secrets would be inconvenient
- to provide an interface to simplify allocation of new securityContexts and
secrets.
These features are explained later.
### Names
From the standpoint of the Kubernetes API, a `user` is any principal which can authenticate to Kubernetes API.
This includes a human running `kubectl` on her desktop and a container in a Pod on a Node making API calls.
From the standpoint of the Kubernetes API, a `user` is any principal which can
authenticate to Kubernetes API. This includes a human running `kubectl` on her
desktop and a container in a Pod on a Node making API calls.
There is already a notion of a username in Kubernetes, which is populated into a request context after authentication.
However, there is no API object representing a user. While this may evolve, it is expected that in mature installations,
the canonical storage of user identifiers will be handled by a system external to Kubernetes.
There is already a notion of a username in Kubernetes, which is populated into a
request context after authentication. However, there is no API object
representing a user. While this may evolve, it is expected that in mature
installations, the canonical storage of user identifiers will be handled by a
system external to Kubernetes.
Kubernetes does not dictate how to divide up the space of user identifier strings. User names can be
simple Unix-style short usernames, (e.g. `alice`), or may be qualified to allow for federated identity (
`alice@example.com` vs `alice@example.org`.) Naming convention may distinguish service accounts from user
accounts (e.g. `alice@example.com` vs `build-service-account-a3b7f0@foo-namespace.service-accounts.example.com`),
but Kubernetes does not require this.
Kubernetes does not dictate how to divide up the space of user identifier
strings. User names can be simple Unix-style short usernames, (e.g. `alice`), or
may be qualified to allow for federated identity (`alice@example.com` vs
`alice@example.org`.) Naming convention may distinguish service accounts from
user accounts (e.g. `alice@example.com` vs
`build-service-account-a3b7f0@foo-namespace.service-accounts.example.com`), but
Kubernetes does not require this.
Kubernetes also does not require that there be a distinction between human and Pod users. It will be possible
to setup a cluster where Alice the human talks to the Kubernetes API as username `alice` and starts pods that
also talk to the API as user `alice` and write files to NFS as user `alice`. But, this is not recommended.
Kubernetes also does not require that there be a distinction between human and
Pod users. It will be possible to setup a cluster where Alice the human talks to
the Kubernetes API as username `alice` and starts pods that also talk to the API
as user `alice` and write files to NFS as user `alice`. But, this is not
recommended.
Instead, it is recommended that Pods and Humans have distinct identities, and reference implementations will
make this distinction.
Instead, it is recommended that Pods and Humans have distinct identities, and
reference implementations will make this distinction.
The distinction is useful for a number of reasons:
- the requirements for humans and automated processes are different:
- Humans need a wide range of capabilities to do their daily activities. Automated processes often have more narrowly-defined activities.
- Humans may better tolerate the exceptional conditions created by expiration of a token. Remembering to handle
this in a program is more annoying. So, either long-lasting credentials or automated rotation of credentials is
needed.
- A Human typically keeps credentials on a machine that is not part of the cluster and so not subject to automatic
management. A VM with a role/service-account can have its credentials automatically managed.
- Humans need a wide range of capabilities to do their daily activities.
Automated processes often have more narrowly-defined activities.
- Humans may better tolerate the exceptional conditions created by
expiration of a token. Remembering to handle this in a program is more annoying.
So, either long-lasting credentials or automated rotation of credentials is
needed.
- A Human typically keeps credentials on a machine that is not part of the
cluster and so not subject to automatic management. A VM with a
role/service-account can have its credentials automatically managed.
- the identity of a Pod cannot in general be mapped to a single human.
- If policy allows, it may be created by one human, and then updated by another, and another, until its behavior cannot be attributed to a single human.
- If policy allows, it may be created by one human, and then updated by
another, and another, until its behavior cannot be attributed to a single human.
**TODO**: consider getting rid of separate serviceAccount object and just rolling its parts into the SecurityContext or
Pod Object.
**TODO**: consider getting rid of separate serviceAccount object and just
rolling its parts into the SecurityContext or Pod Object.
The `secrets` field is a list of references to /secret objects that an process started as that service account should
have access to be able to assert that role.
The `secrets` field is a list of references to /secret objects that an process
started as that service account should have access to be able to assert that
role.
The secrets are not inline with the serviceAccount object. This way, most or all users can have permission to `GET /serviceAccounts` so they can remind themselves
what serviceAccounts are available for use.
The secrets are not inline with the serviceAccount object. This way, most or
all users can have permission to `GET /serviceAccounts` so they can remind
themselves what serviceAccounts are available for use.
Nothing will prevent creation of a serviceAccount with two secrets of type `SecretTypeKubernetesAuth`, or secrets of two
different types. Kubelet and client libraries will have some behavior, TBD, to handle the case of multiple secrets of a
given type (pick first or provide all and try each in order, etc).
Nothing will prevent creation of a serviceAccount with two secrets of type
`SecretTypeKubernetesAuth`, or secrets of two different types. Kubelet and
client libraries will have some behavior, TBD, to handle the case of multiple
secrets of a given type (pick first or provide all and try each in order, etc).
When a serviceAccount and a matching secret exist, then a `User.Info` for the serviceAccount and a `BearerToken` from the secret
are added to the map of tokens used by the authentication process in the apiserver, and similarly for other types. (We
might have some types that do not do anything on apiserver but just get pushed to the kubelet.)
When a serviceAccount and a matching secret exist, then a `User.Info` for the
serviceAccount and a `BearerToken` from the secret are added to the map of
tokens used by the authentication process in the apiserver, and similarly for
other types. (We might have some types that do not do anything on apiserver but
just get pushed to the kubelet.)
### Pods
The `PodSpec` is extended to have a `Pods.Spec.ServiceAccountUsername` field. If this is unset, then a
default value is chosen. If it is set, then the corresponding value of `Pods.Spec.SecurityContext` is set by the
Service Account Finalizer (see below).
The `PodSpec` is extended to have a `Pods.Spec.ServiceAccountUsername` field. If
this is unset, then a default value is chosen. If it is set, then the
corresponding value of `Pods.Spec.SecurityContext` is set by the Service Account
Finalizer (see below).
TBD: how policy limits which users can make pods with which service accounts.
### Authorization
Kubernetes API Authorization Policies refer to users. Pods created with a `Pods.Spec.ServiceAccountUsername` typically
get a `Secret` which allows them to authenticate to the Kubernetes APIserver as a particular user. So any
policy that is desired can be applied to them.
Kubernetes API Authorization Policies refer to users. Pods created with a
`Pods.Spec.ServiceAccountUsername` typically get a `Secret` which allows them to
authenticate to the Kubernetes APIserver as a particular user. So any policy
that is desired can be applied to them.
A higher level workflow is needed to coordinate creation of serviceAccounts, secrets and relevant policy objects.
Users are free to extend Kubernetes to put this business logic wherever is convenient for them, though the
Service Account Finalizer is one place where this can happen (see below).
A higher level workflow is needed to coordinate creation of serviceAccounts,
secrets and relevant policy objects. Users are free to extend Kubernetes to put
this business logic wherever is convenient for them, though the Service Account
Finalizer is one place where this can happen (see below).
### Kubelet
The kubelet will treat as "not ready to run" (needing a finalizer to act on it) any Pod which has an empty
SecurityContext.
The kubelet will treat as "not ready to run" (needing a finalizer to act on it)
any Pod which has an empty SecurityContext.
The kubelet will set a default, restrictive, security context for any pods created from non-Apiserver config
sources (http, file).
The kubelet will set a default, restrictive, security context for any pods
created from non-Apiserver config sources (http, file).
Kubelet watches apiserver for secrets which are needed by pods bound to it.
@ -173,32 +202,41 @@ Kubelet watches apiserver for secrets which are needed by pods bound to it.
There are several ways to use Pods with SecurityContexts and Secrets.
One way is to explicitly specify the securityContext and all secrets of a Pod when the pod is initially created,
like this:
One way is to explicitly specify the securityContext and all secrets of a Pod
when the pod is initially created, like this:
**TODO**: example of pod with explicit refs.
Another way is with the *Service Account Finalizer*, a plugin process which is optional, and which handles
business logic around service accounts.
Another way is with the *Service Account Finalizer*, a plugin process which is
optional, and which handles business logic around service accounts.
The Service Account Finalizer watches Pods, Namespaces, and ServiceAccount definitions.
The Service Account Finalizer watches Pods, Namespaces, and ServiceAccount
definitions.
First, if it finds pods which have a `Pod.Spec.ServiceAccountUsername` but no `Pod.Spec.SecurityContext` set,
then it copies in the referenced securityContext and secrets references for the corresponding `serviceAccount`.
First, if it finds pods which have a `Pod.Spec.ServiceAccountUsername` but no
`Pod.Spec.SecurityContext` set, then it copies in the referenced securityContext
and secrets references for the corresponding `serviceAccount`.
Second, if ServiceAccount definitions change, it may take some actions.
**TODO**: decide what actions it takes when a serviceAccount definition changes. Does it stop pods, or just
allow someone to list ones that are out of spec? In general, people may want to customize this?
Third, if a new namespace is created, it may create a new serviceAccount for that namespace. This may include
a new username (e.g. `NAMESPACE-default-service-account@serviceaccounts.$CLUSTERID.kubernetes.io`), a new
securityContext, a newly generated secret to authenticate that serviceAccount to the Kubernetes API, and default
policies for that service account.
**TODO**: more concrete example. What are typical default permissions for default service account (e.g. readonly access
to services in the same namespace and read-write access to events in that namespace?)
**TODO**: decide what actions it takes when a serviceAccount definition changes.
Does it stop pods, or just allow someone to list ones that are out of spec? In
general, people may want to customize this?
Finally, it may provide an interface to automate creation of new serviceAccounts. In that case, the user may want
to GET serviceAccounts to see what has been created.
Third, if a new namespace is created, it may create a new serviceAccount for
that namespace. This may include a new username (e.g.
`NAMESPACE-default-service-account@serviceaccounts.$CLUSTERID.kubernetes.io`),
a new securityContext, a newly generated secret to authenticate that
serviceAccount to the Kubernetes API, and default policies for that service
account.
**TODO**: more concrete example. What are typical default permissions for
default service account (e.g. readonly access to services in the same namespace
and read-write access to events in that namespace?)
Finally, it may provide an interface to automate creation of new
serviceAccounts. In that case, the user may want to GET serviceAccounts to see
what has been created.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->

77
docs/design/simple-rolling-update.md
View File

@ -34,32 +34,47 @@ Documentation for other releases can be found at
## Simple rolling update
This is a lightweight design document for simple [rolling update](../user-guide/kubectl/kubectl_rolling-update.md) in `kubectl`.
This is a lightweight design document for simple
[rolling update](../user-guide/kubectl/kubectl_rolling-update.md) in `kubectl`.
Complete execution flow can be found [here](#execution-details). See the [example of rolling update](../user-guide/update-demo/) for more information.
Complete execution flow can be found [here](#execution-details). See the
[example of rolling update](../user-guide/update-demo/) for more information.
### Lightweight rollout
Assume that we have a current replication controller named `foo` and it is running image `image:v1`
Assume that we have a current replication controller named `foo` and it is
running image `image:v1`
`kubectl rolling-update foo [foo-v2] --image=myimage:v2`
If the user doesn't specify a name for the 'next' replication controller, then the 'next' replication controller is renamed to
If the user doesn't specify a name for the 'next' replication controller, then
the 'next' replication controller is renamed to
the name of the original replication controller.
Obviously there is a race here, where if you kill the client between delete foo, and creating the new version of 'foo' you might be surprised about what is there, but I think that's ok.
See [Recovery](#recovery) below
Obviously there is a race here, where if you kill the client between delete foo,
and creating the new version of 'foo' you might be surprised about what is
there, but I think that's ok. See [Recovery](#recovery) below
If the user does specify a name for the 'next' replication controller, then the 'next' replication controller is retained with its existing name,
and the old 'foo' replication controller is deleted. For the purposes of the rollout, we add a unique-ifying label `kubernetes.io/deployment` to both the `foo` and `foo-next` replication controllers.
The value of that label is the hash of the complete JSON representation of the`foo-next` or`foo` replication controller. The name of this label can be overridden by the user with the `--deployment-label-key` flag.
If the user does specify a name for the 'next' replication controller, then the
'next' replication controller is retained with its existing name, and the old
'foo' replication controller is deleted. For the purposes of the rollout, we add
a unique-ifying label `kubernetes.io/deployment` to both the `foo` and
`foo-next` replication controllers. The value of that label is the hash of the
complete JSON representation of the`foo-next` or`foo` replication controller.
The name of this label can be overridden by the user with the
`--deployment-label-key` flag.
#### Recovery
If a rollout fails or is terminated in the middle, it is important that the user be able to resume the roll out.
To facilitate recovery in the case of a crash of the updating process itself, we add the following annotations to each replication controller in the `kubernetes.io/` annotation namespace:
* `desired-replicas` The desired number of replicas for this replication controller (either N or zero)
* `update-partner` A pointer to the replication controller resource that is the other half of this update (syntax `<name>` the namespace is assumed to be identical to the namespace of this replication controller.)
If a rollout fails or is terminated in the middle, it is important that the user
be able to resume the roll out. To facilitate recovery in the case of a crash of
the updating process itself, we add the following annotations to each
replication controller in the `kubernetes.io/` annotation namespace:
* `desired-replicas` The desired number of replicas for this replication
controller (either N or zero)
* `update-partner` A pointer to the replication controller resource that is
the other half of this update (syntax `<name>` the namespace is assumed to be
identical to the namespace of this replication controller.)
Recovery is achieved by issuing the same command again:
@ -67,9 +82,12 @@ Recovery is achieved by issuing the same command again:
kubectl rolling-update foo [foo-v2] --image=myimage:v2
```
Whenever the rolling update command executes, the kubectl client looks for replication controllers called `foo` and `foo-next`, if they exist, an attempt is
made to roll `foo` to `foo-next`. If `foo-next` does not exist, then it is created, and the rollout is a new rollout. If `foo` doesn't exist, then
it is assumed that the rollout is nearly completed, and `foo-next` is renamed to `foo`. Details of the execution flow are given below.
Whenever the rolling update command executes, the kubectl client looks for
replication controllers called `foo` and `foo-next`, if they exist, an attempt
is made to roll `foo` to `foo-next`. If `foo-next` does not exist, then it is
created, and the rollout is a new rollout. If `foo` doesn't exist, then it is
assumed that the rollout is nearly completed, and `foo-next` is renamed to
`foo`. Details of the execution flow are given below.
### Aborting a rollout
@ -82,22 +100,28 @@ This is really just semantic sugar for:
`kubectl rolling-update foo-v2 foo`
With the added detail that it moves the `desired-replicas` annotation from `foo-v2` to `foo`
With the added detail that it moves the `desired-replicas` annotation from
`foo-v2` to `foo`
### Execution Details
For the purposes of this example, assume that we are rolling from `foo` to `foo-next` where the only change is an image update from `v1` to `v2`
For the purposes of this example, assume that we are rolling from `foo` to
`foo-next` where the only change is an image update from `v1` to `v2`
If the user doesn't specify a `foo-next` name, then it is either discovered from the `update-partner` annotation on `foo`. If that annotation doesn't exist,
then `foo-next` is synthesized using the pattern `<controller-name>-<hash-of-next-controller-JSON>`
If the user doesn't specify a `foo-next` name, then it is either discovered from
the `update-partner` annotation on `foo`. If that annotation doesn't exist,
then `foo-next` is synthesized using the pattern
`<controller-name>-<hash-of-next-controller-JSON>`
#### Initialization
* If `foo` and `foo-next` do not exist:
* Exit, and indicate an error to the user, that the specified controller doesn't exist.
* Exit, and indicate an error to the user, that the specified controller
doesn't exist.
* If `foo` exists, but `foo-next` does not:
* Create `foo-next` populate it with the `v2` image, set `desired-replicas` to `foo.Spec.Replicas`
* Create `foo-next` populate it with the `v2` image, set
`desired-replicas` to `foo.Spec.Replicas`
* Goto Rollout
* If `foo-next` exists, but `foo` does not:
* Assume that we are in the rename phase.
@ -105,7 +129,8 @@ then `foo-next` is synthesized using the pattern `<controller-name>-<hash-of-nex
* If both `foo` and `foo-next` exist:
* Assume that we are in a partial rollout
* If `foo-next` is missing the `desired-replicas` annotation
* Populate the `desired-replicas` annotation to `foo-next` using the current size of `foo`
* Populate the `desired-replicas` annotation to `foo-next` using the
current size of `foo`
* Goto Rollout
#### Rollout
@ -125,11 +150,13 @@ then `foo-next` is synthesized using the pattern `<controller-name>-<hash-of-nex
#### Abort
* If `foo-next` doesn't exist
* Exit and indicate to the user that they may want to simply do a new rollout with the old version
* Exit and indicate to the user that they may want to simply do a new
rollout with the old version
* If `foo` doesn't exist
* Exit and indicate not found to the user
* Otherwise, `foo-next` and `foo` both exist
* Set `desired-replicas` annotation on `foo` to match the annotation on `foo-next`
* Set `desired-replicas` annotation on `foo` to match the annotation on
`foo-next`
* Goto Rollout with `foo` and `foo-next` trading places.

298
docs/design/taint-toleration-dedicated.md
View File

@ -36,75 +36,85 @@ Documentation for other releases can be found at
## Introduction
This document describes *taints* and *tolerations*, which constitute a generic mechanism for restricting
the set of pods that can use a node. We also describe one concrete use case for the mechanism,
namely to limit the set of users (or more generally, authorization domains)
who can access a set of nodes (a feature we call
*dedicated nodes*). There are many other uses--for example, a set of nodes with a particular
piece of hardware could
be reserved for pods that require that hardware, or a node could be marked as unschedulable
when it is being drained before shutdown, or a node could trigger evictions when it experiences
hardware or software problems or abnormal node configurations; see #17190 and #3885 for more discussion.
This document describes *taints* and *tolerations*, which constitute a generic
mechanism for restricting the set of pods that can use a node. We also describe
one concrete use case for the mechanism, namely to limit the set of users (or
more generally, authorization domains) who can access a set of nodes (a feature
we call *dedicated nodes*). There are many other uses--for example, a set of
nodes with a particular piece of hardware could be reserved for pods that
require that hardware, or a node could be marked as unschedulable when it is
being drained before shutdown, or a node could trigger evictions when it
experiences hardware or software problems or abnormal node configurations; see
issues #17190 and #3885 for more discussion.
## Taints, tolerations, and dedicated nodes
A *taint* is a new type that is part of the `NodeSpec`; when present, it prevents pods
from scheduling onto the node unless the pod *tolerates* the taint (tolerations are listed
in the `PodSpec`). Note that there are actually multiple flavors of taints: taints that
prevent scheduling on a node, taints that cause the scheduler to try to avoid scheduling
on a node but do not prevent it, taints that prevent a pod from starting on Kubelet even
if the pod's `NodeName` was written directly (i.e. pod did not go through the scheduler),
and taints that evict already-running pods.
A *taint* is a new type that is part of the `NodeSpec`; when present, it
prevents pods from scheduling onto the node unless the pod *tolerates* the taint
(tolerations are listed in the `PodSpec`). Note that there are actually multiple
flavors of taints: taints that prevent scheduling on a node, taints that cause
the scheduler to try to avoid scheduling on a node but do not prevent it, taints
that prevent a pod from starting on Kubelet even if the pod's `NodeName` was
written directly (i.e. pod did not go through the scheduler), and taints that
evict already-running pods.
[This comment](https://github.com/kubernetes/kubernetes/issues/3885#issuecomment-146002375)
has more background on these different scenarios. We will focus on the first
kind of taint in this doc, since it is the kind required for the "dedicated nodes" use case.
kind of taint in this doc, since it is the kind required for the "dedicated
nodes" use case.
Implementing dedicated nodes using taints and tolerations is straightforward: in essence, a node that
is dedicated to group A gets taint `dedicated=A` and the pods belonging to group A get
toleration `dedicated=A`. (The exact syntax and semantics of taints and tolerations are
described later in this doc.) This keeps all pods except those belonging to group A off of the nodes.
This approach easily generalizes to pods that are allowed to
schedule into multiple dedicated node groups, and nodes that are a member of multiple
dedicated node groups.
Implementing dedicated nodes using taints and tolerations is straightforward: in
essence, a node that is dedicated to group A gets taint `dedicated=A` and the
pods belonging to group A get toleration `dedicated=A`. (The exact syntax and
semantics of taints and tolerations are described later in this doc.) This keeps
all pods except those belonging to group A off of the nodes. This approach
easily generalizes to pods that are allowed to schedule into multiple dedicated
node groups, and nodes that are a member of multiple dedicated node groups.
Note that because tolerations are at the granularity of pods,
the mechanism is very flexible -- any policy can be used to determine which tolerations
should be placed on a pod. So the "group A" mentioned above could be all pods from a
particular namespace or set of namespaces, or all pods with some other arbitrary characteristic
in common. We expect that any real-world usage of taints and tolerations will employ an admission controller
to apply the tolerations. For example, to give all pods from namespace A access to dedicated
node group A, an admission controller would add the corresponding toleration to all
pods from namespace A. Or to give all pods that require GPUs access to GPU nodes, an admission
controller would add the toleration for GPU taints to pods that request the GPU resource.
Note that because tolerations are at the granularity of pods, the mechanism is
very flexible -- any policy can be used to determine which tolerations should be
placed on a pod. So the "group A" mentioned above could be all pods from a
particular namespace or set of namespaces, or all pods with some other arbitrary
characteristic in common. We expect that any real-world usage of taints and
tolerations will employ an admission controller to apply the tolerations. For
example, to give all pods from namespace A access to dedicated node group A, an
admission controller would add the corresponding toleration to all pods from
namespace A. Or to give all pods that require GPUs access to GPU nodes, an
admission controller would add the toleration for GPU taints to pods that
request the GPU resource.
Everything that can be expressed using taints and tolerations can be expressed using
[node affinity](https://github.com/kubernetes/kubernetes/pull/18261), e.g. in the example
in the previous paragraph, you could put a label `dedicated=A` on the set of dedicated nodes and
a node affinity `dedicated NotIn A` on all pods *not* belonging to group A. But it is
cumbersome to express exclusion policies using node affinity because every time you add
a new type of restricted node, all pods that aren't allowed to use those nodes need to start avoiding those
nodes using node affinity. This means the node affinity list can get quite long in clusters with lots of different
groups of special nodes (lots of dedicated node groups, lots of different kinds of special hardware, etc.).
Moreover, you need to also update any Pending pods when you add new types of special nodes.
In contrast, with taints and tolerations,
when you add a new type of special node, "regular" pods are unaffected, and you just need to add
the necessary toleration to the pods you subsequent create that need to use the new type of special nodes.
To put it another way, with taints and tolerations, only pods that use a set of special nodes
need to know about those special nodes; with the node affinity approach, pods that have
no interest in those special nodes need to know about all of the groups of special nodes.
Everything that can be expressed using taints and tolerations can be expressed
using [node affinity](https://github.com/kubernetes/kubernetes/pull/18261), e.g.
in the example in the previous paragraph, you could put a label `dedicated=A` on
the set of dedicated nodes and a node affinity `dedicated NotIn A` on all pods *not*
belonging to group A. But it is cumbersome to express exclusion policies using
node affinity because every time you add a new type of restricted node, all pods
that aren't allowed to use those nodes need to start avoiding those nodes using
node affinity. This means the node affinity list can get quite long in clusters
with lots of different groups of special nodes (lots of dedicated node groups,
lots of different kinds of special hardware, etc.). Moreover, you need to also
update any Pending pods when you add new types of special nodes. In contrast,
with taints and tolerations, when you add a new type of special node, "regular"
pods are unaffected, and you just need to add the necessary toleration to the
pods you subsequent create that need to use the new type of special nodes. To
put it another way, with taints and tolerations, only pods that use a set of
special nodes need to know about those special nodes; with the node affinity
approach, pods that have no interest in those special nodes need to know about
all of the groups of special nodes.
One final comment: in practice, it is often desirable to not
only keep "regular" pods off of special nodes, but also to keep "special" pods off of
regular nodes. An example in the dedicated nodes case is to not only keep regular
users off of dedicated nodes, but also to keep dedicated users off of non-dedicated (shared)
nodes. In this case, the "non-dedicated" nodes can be modeled as their own dedicated node group
(for example, tainted as `dedicated=shared`), and pods that are not given access to any
dedicated nodes ("regular" pods) would be given a toleration for `dedicated=shared`. (As mentioned earlier,
we expect tolerations will be added by an admission controller.) In this case taints/tolerations
are still better than node affinity because with taints/tolerations each pod only needs one special "marking",
versus in the node affinity case where every time you add a dedicated node group (i.e. a new
`dedicated=` value), you need to add a new node affinity rule to all pods (including pending pods)
except the ones allowed to use that new dedicated node group.
One final comment: in practice, it is often desirable to not only keep "regular"
pods off of special nodes, but also to keep "special" pods off of regular nodes.
An example in the dedicated nodes case is to not only keep regular users off of
dedicated nodes, but also to keep dedicated users off of non-dedicated (shared)
nodes. In this case, the "non-dedicated" nodes can be modeled as their own
dedicated node group (for example, tainted as `dedicated=shared`), and pods that
are not given access to any dedicated nodes ("regular" pods) would be given a
toleration for `dedicated=shared`. (As mentioned earlier, we expect tolerations
will be added by an admission controller.) In this case taints/tolerations are
still better than node affinity because with taints/tolerations each pod only
needs one special "marking", versus in the node affinity case where every time
you add a dedicated node group (i.e. a new `dedicated=` value), you need to add
a new node affinity rule to all pods (including pending pods) except the ones
allowed to use that new dedicated node group.
## API
@ -112,56 +122,56 @@ except the ones allowed to use that new dedicated node group.
// The node this Taint is attached to has the effect "effect" on
// any pod that that does not tolerate the Taint.
type Taint struct {
Key string `json:"key" patchStrategy:"merge" patchMergeKey:"key"`
Value string `json:"value,omitempty"`
Effect TaintEffect `json:"effect"`
Key string `json:"key" patchStrategy:"merge" patchMergeKey:"key"`
Value string `json:"value,omitempty"`
Effect TaintEffect `json:"effect"`
}
type TaintEffect string
const (
// Do not allow new pods to schedule unless they tolerate the taint,
// but allow all pods submitted to Kubelet without going through the scheduler
// to start, and allow all already-running pods to continue running.
// Enforced by the scheduler.
TaintEffectNoSchedule TaintEffect = "NoSchedule"
// Like TaintEffectNoSchedule, but the scheduler tries not to schedule
// new pods onto the node, rather than prohibiting new pods from scheduling
// onto the node. Enforced by the scheduler.
TaintEffectPreferNoSchedule TaintEffect = "PreferNoSchedule"
// Do not allow new pods to schedule unless they tolerate the taint,
// do not allow pods to start on Kubelet unless they tolerate the taint,
// but allow all already-running pods to continue running.
// Enforced by the scheduler and Kubelet.
TaintEffectNoScheduleNoAdmit TaintEffect = "NoScheduleNoAdmit"
// Do not allow new pods to schedule unless they tolerate the taint,
// do not allow pods to start on Kubelet unless they tolerate the taint,
// and try to eventually evict any already-running pods that do not tolerate the taint.
// Enforced by the scheduler and Kubelet.
TaintEffectNoScheduleNoAdmitNoExecute = "NoScheduleNoAdmitNoExecute"
// Do not allow new pods to schedule unless they tolerate the taint,
// but allow all pods submitted to Kubelet without going through the scheduler
// to start, and allow all already-running pods to continue running.
// Enforced by the scheduler.
TaintEffectNoSchedule TaintEffect = "NoSchedule"
// Like TaintEffectNoSchedule, but the scheduler tries not to schedule
// new pods onto the node, rather than prohibiting new pods from scheduling
// onto the node. Enforced by the scheduler.
TaintEffectPreferNoSchedule TaintEffect = "PreferNoSchedule"
// Do not allow new pods to schedule unless they tolerate the taint,
// do not allow pods to start on Kubelet unless they tolerate the taint,
// but allow all already-running pods to continue running.
// Enforced by the scheduler and Kubelet.
TaintEffectNoScheduleNoAdmit TaintEffect = "NoScheduleNoAdmit"
// Do not allow new pods to schedule unless they tolerate the taint,
// do not allow pods to start on Kubelet unless they tolerate the taint,
// and try to eventually evict any already-running pods that do not tolerate the taint.
// Enforced by the scheduler and Kubelet.
TaintEffectNoScheduleNoAdmitNoExecute = "NoScheduleNoAdmitNoExecute"
)
// The pod this Toleration is attached to tolerates any taint that matches
// the triple <key,value,effect> using the matching operator <operator>.
type Toleration struct {
Key string `json:"key" patchStrategy:"merge" patchMergeKey:"key"`
// operator represents a key's relationship to the value.
// Valid operators are Exists and Equal. Defaults to Equal.
// Exists is equivalent to wildcard for value, so that a pod can
// tolerate all taints of a particular category.
Operator TolerationOperator `json:"operator"`
Value string `json:"value,omitempty"`
Effect TaintEffect `json:"effect"`
// TODO: For forgiveness (#1574), we'd eventually add at least a grace period
// here, and possibly an occurrence threshold and period.
Key string `json:"key" patchStrategy:"merge" patchMergeKey:"key"`
// operator represents a key's relationship to the value.
// Valid operators are Exists and Equal. Defaults to Equal.
// Exists is equivalent to wildcard for value, so that a pod can
// tolerate all taints of a particular category.
Operator TolerationOperator `json:"operator"`
Value string `json:"value,omitempty"`
Effect TaintEffect `json:"effect"`
// TODO: For forgiveness (#1574), we'd eventually add at least a grace period
// here, and possibly an occurrence threshold and period.
}
// A toleration operator is the set of operators that can be used in a toleration.
type TolerationOperator string
const (
TolerationOpExists TolerationOperator = "Exists"
TolerationOpEqual TolerationOperator = "Equal"
TolerationOpExists TolerationOperator = "Exists"
TolerationOpEqual TolerationOperator = "Equal"
)
```
@ -169,18 +179,17 @@ const (
(See [this comment](https://github.com/kubernetes/kubernetes/issues/3885#issuecomment-146002375)
to understand the motivation for the various taint effects.)
We will add
We will add:
```go
// Multiple tolerations with the same key are allowed.
Tolerations []Toleration `json:"tolerations,omitempty"`
```
to `PodSpec`. A pod must tolerate all of a node's taints (except taints
of type TaintEffectPreferNoSchedule) in order to be able
to schedule onto that node.
to `PodSpec`. A pod must tolerate all of a node's taints (except taints of type
TaintEffectPreferNoSchedule) in order to be able to schedule onto that node.
We will add
We will add:
```go
// Multiple taints with the same key are not allowed.
@ -201,30 +210,32 @@ Taints and tolerations are not scoped to namespace.
Using taints and tolerations to implement dedicated nodes requires these steps:
1. Add the API described above
1. Add a scheduler predicate function that respects taints and tolerations (for TaintEffectNoSchedule)
and a scheduler priority function that respects taints and tolerations (for TaintEffectPreferNoSchedule).
1. Add to the Kubelet code to implement the "no admit" behavior of TaintEffectNoScheduleNoAdmit and
TaintEffectNoScheduleNoAdmitNoExecute
1. Add a scheduler predicate function that respects taints and tolerations (for
TaintEffectNoSchedule) and a scheduler priority function that respects taints
and tolerations (for TaintEffectPreferNoSchedule).
1. Add to the Kubelet code to implement the "no admit" behavior of
TaintEffectNoScheduleNoAdmit and TaintEffectNoScheduleNoAdmitNoExecute
1. Implement code in Kubelet that evicts a pod that no longer satisfies
TaintEffectNoScheduleNoAdmitNoExecute. In theory we could do this in the controllers
instead, but since taints might be used to enforce security policies, it is better
to do in kubelet because kubelet can respond quickly and can guarantee the rules will
be applied to all pods.
Eviction may need to happen under a variety of circumstances: when a taint is added, when an existing
taint is updated, when a toleration is removed from a pod, or when a toleration is modified on a pod.
TaintEffectNoScheduleNoAdmitNoExecute. In theory we could do this in the
controllers instead, but since taints might be used to enforce security
policies, it is better to do in kubelet because kubelet can respond quickly and
can guarantee the rules will be applied to all pods. Eviction may need to happen
under a variety of circumstances: when a taint is added, when an existing taint
is updated, when a toleration is removed from a pod, or when a toleration is
modified on a pod.
1. Add a new `kubectl` command that adds/removes taints to/from nodes,
1. (This is the one step is that is specific to dedicated nodes)
Implement an admission controller that adds tolerations to pods that are supposed
to be allowed to use dedicated nodes (for example, based on pod's namespace).
1. (This is the one step is that is specific to dedicated nodes) Implement an
admission controller that adds tolerations to pods that are supposed to be
allowed to use dedicated nodes (for example, based on pod's namespace).
In the future one can imagine a generic policy configuration that configures
an admission controller to apply the appropriate tolerations to the desired class of pods and
taints to Nodes upon node creation. It could be used not just for policies about dedicated nodes,
but also other uses of taints and tolerations, e.g. nodes that are restricted
due to their hardware configuration.
In the future one can imagine a generic policy configuration that configures an
admission controller to apply the appropriate tolerations to the desired class
of pods and taints to Nodes upon node creation. It could be used not just for
policies about dedicated nodes, but also other uses of taints and tolerations,
e.g. nodes that are restricted due to their hardware configuration.
The `kubectl` command to add and remove taints on nodes will be modeled after `kubectl label`.
Examples usages:
The `kubectl` command to add and remove taints on nodes will be modeled after
`kubectl label`. Examples usages:
```sh
# Update node 'foo' with a taint with key 'dedicated' and value 'special-user' and effect 'NoScheduleNoAdmitNoExecute'.
@ -258,36 +269,41 @@ to enumerate them by name.
## Future work
At present, the Kubernetes security model allows any user to add and remove any taints and tolerations.
Obviously this makes it impossible to securely enforce
rules like dedicated nodes. We need some mechanism that prevents regular users from mutating the `Taints`
field of `NodeSpec` (probably we want to prevent them from mutating any fields of `NodeSpec`)
and from mutating the `Tolerations` field of their pods. #17549 is relevant.
At present, the Kubernetes security model allows any user to add and remove any
taints and tolerations. Obviously this makes it impossible to securely enforce
rules like dedicated nodes. We need some mechanism that prevents regular users
from mutating the `Taints` field of `NodeSpec` (probably we want to prevent them
from mutating any fields of `NodeSpec`) and from mutating the `Tolerations`
field of their pods. #17549 is relevant.
Another security vulnterability arises if nodes are added to the cluster before receiving
their taint. Thus we need to ensure that a new node does not become "Ready" until it has been
configured with its taints. One way to do this is to have an admission controller that adds the taint whenever
a Node object is created.
Another security vulnerability arises if nodes are added to the cluster before
receiving their taint. Thus we need to ensure that a new node does not become
"Ready" until it has been configured with its taints. One way to do this is to
have an admission controller that adds the taint whenever a Node object is
created.
A quota policy may want to treat nodes differently based on what taints, if any,
they have. For example, if a particular namespace is only allowed to access dedicated nodes,
then it may be convenient to give the namespace unlimited quota. (To use finite quota,
you'd have to size the namespace's quota to the sum of the sizes of the machines in the
dedicated node group, and update it when nodes are added/removed to/from the group.)
they have. For example, if a particular namespace is only allowed to access
dedicated nodes, then it may be convenient to give the namespace unlimited
quota. (To use finite quota, you'd have to size the namespace's quota to the sum
of the sizes of the machines in the dedicated node group, and update it when
nodes are added/removed to/from the group.)
It's conceivable that taints and tolerations could be unified with [pod anti-affinity](https://github.com/kubernetes/kubernetes/pull/18265).
We have chosen not to do this for the reasons described in the "Future work" section of that doc.
It's conceivable that taints and tolerations could be unified with
[pod anti-affinity](https://github.com/kubernetes/kubernetes/pull/18265).
We have chosen not to do this for the reasons described in the "Future work"
section of that doc.
## Backward compatibility
Old scheduler versions will ignore taints and tolerations. New scheduler versions
will respect them.
Old scheduler versions will ignore taints and tolerations. New scheduler
versions will respect them.
Users should not start using taints and tolerations until the full implementation
has been in Kubelet and the master for enough binary versions that we
feel comfortable that we will not need to roll back either Kubelet or
master to a version that does not support them. Longer-term we will
use a progamatic approach to enforcing this (#4855).
Users should not start using taints and tolerations until the full
implementation has been in Kubelet and the master for enough binary versions
that we feel comfortable that we will not need to roll back either Kubelet or
master to a version that does not support them. Longer-term we will use a
progamatic approach to enforcing this (#4855).
## Related issues

134
docs/design/versioning.md
View File

@ -38,7 +38,9 @@ Reference: [Semantic Versioning](http://semver.org)
Legend:
* **Kube X.Y.Z** refers to the version (git tag) of Kubernetes that is released. This versions all components: apiserver, kubelet, kubectl, etc. (**X** is the major version, **Y** is the minor version, and **Z** is the patch version.)
* **Kube X.Y.Z** refers to the version (git tag) of Kubernetes that is released.
This versions all components: apiserver, kubelet, kubectl, etc. (**X** is the
major version, **Y** is the minor version, and **Z** is the patch version.)
* **API vX[betaY]** refers to the version of the HTTP API.
## Release versioning
@ -46,43 +48,76 @@ Legend:
### Minor version scheme and timeline
* Kube X.Y.0-alpha.W, W > 0 (Branch: master)
* Alpha releases are released roughly every two weeks directly from the master branch.
* No cherrypick releases. If there is a critical bugfix, a new release from master can be created ahead of schedule.
* Alpha releases are released roughly every two weeks directly from the master
branch.
* No cherrypick releases. If there is a critical bugfix, a new release from
master can be created ahead of schedule.
* Kube X.Y.Z-beta.W (Branch: release-X.Y)
* When master is feature-complete for Kube X.Y, we will cut the release-X.Y branch 2 weeks prior to the desired X.Y.0 date and cherrypick only PRs essential to X.Y.
* When master is feature-complete for Kube X.Y, we will cut the release-X.Y
branch 2 weeks prior to the desired X.Y.0 date and cherrypick only PRs essential
to X.Y.
* This cut will be marked as X.Y.0-beta.0, and master will be revved to X.Y+1.0-alpha.0.
* If we're not satisfied with X.Y.0-beta.0, we'll release other beta releases, (X.Y.0-beta.W | W > 0) as necessary.
* If we're not satisfied with X.Y.0-beta.0, we'll release other beta releases,
(X.Y.0-beta.W | W > 0) as necessary.
* Kube X.Y.0 (Branch: release-X.Y)
* Final release, cut from the release-X.Y branch cut two weeks prior.
* X.Y.1-beta.0 will be tagged at the same commit on the same branch.
* X.Y.0 occur 3 to 4 months after X.(Y-1).0.
* Kube X.Y.Z, Z > 0 (Branch: release-X.Y)
* [Patch releases](#patch-releases) are released as we cherrypick commits into the release-X.Y branch, (which is at X.Y.Z-beta.W,) as needed.
* X.Y.Z is cut straight from the release-X.Y branch, and X.Y.Z+1-beta.0 is tagged on the followup commit that updates pkg/version/base.go with the beta version.
* [Patch releases](#patch-releases) are released as we cherrypick commits into
the release-X.Y branch, (which is at X.Y.Z-beta.W,) as needed.
* X.Y.Z is cut straight from the release-X.Y branch, and X.Y.Z+1-beta.0 is
tagged on the followup commit that updates pkg/version/base.go with the beta
version.
* Kube X.Y.Z, Z > 0 (Branch: release-X.Y.Z)
* These are special and different in that the X.Y.Z tag is branched to isolate the emergency/critical fix from all other changes that have landed on the release branch since the previous tag
* These are special and different in that the X.Y.Z tag is branched to isolate
the emergency/critical fix from all other changes that have landed on the
release branch since the previous tag
* Cut release-X.Y.Z branch to hold the isolated patch release
* Tag release-X.Y.Z branch + fixes with X.Y.(Z+1)
* Branched [patch releases](#patch-releases) are rarely needed but used for emergency/critical fixes to the latest release
* See [#19849](https://issues.k8s.io/19849) tracking the work that is needed for this kind of release to be possible.
* Branched [patch releases](#patch-releases) are rarely needed but used for
emergency/critical fixes to the latest release
* See [#19849](https://issues.k8s.io/19849) tracking the work that is needed
for this kind of release to be possible.
### Major version timeline
There is no mandated timeline for major versions. They only occur when we need to start the clock on deprecating features. A given major version should be the latest major version for at least one year from its original release date.
There is no mandated timeline for major versions. They only occur when we need
to start the clock on deprecating features. A given major version should be the
latest major version for at least one year from its original release date.
### CI and dev version scheme
* Continuous integration versions also exist, and are versioned off of alpha and beta releases. X.Y.Z-alpha.W.C+aaaa is C commits after X.Y.Z-alpha.W, with an additional +aaaa build suffix added; X.Y.Z-beta.W.C+bbbb is C commits after X.Y.Z-beta.W, with an additional +bbbb build suffix added. Furthermore, builds that are built off of a dirty build tree, (during development, with things in the tree that are not checked it,) it will be appended with -dirty.
* Continuous integration versions also exist, and are versioned off of alpha and
beta releases. X.Y.Z-alpha.W.C+aaaa is C commits after X.Y.Z-alpha.W, with an
additional +aaaa build suffix added; X.Y.Z-beta.W.C+bbbb is C commits after
X.Y.Z-beta.W, with an additional +bbbb build suffix added. Furthermore, builds
that are built off of a dirty build tree, (during development, with things in
the tree that are not checked it,) it will be appended with -dirty.
### Supported releases
We expect users to stay reasonably up-to-date with the versions of Kubernetes they use in production, but understand that it may take time to upgrade.
We expect users to stay reasonably up-to-date with the versions of Kubernetes
they use in production, but understand that it may take time to upgrade.
We expect users to be running approximately the latest patch release of a given minor release; we often include critical bug fixes in [patch releases](#patch-release), and so encourage users to upgrade as soon as possible. Furthermore, we expect to "support" three minor releases at a time. "Support" means we expect users to be running that version in production, though we may not port fixes back before the latest minor version. For example, when v1.3 comes out, v1.0 will no longer be supported: basically, that means that the reasonable response to the question "my v1.0 cluster isn't working," is, "you should probably upgrade it, (and probably should have some time ago)". With minor releases happening approximately every three months, that means a minor release is supported for approximately nine months.
We expect users to be running approximately the latest patch release of a given
minor release; we often include critical bug fixes in
[patch releases](#patch-release), and so encourage users to upgrade as soon as
possible. Furthermore, we expect to "support" three minor releases at a time.
"Support" means we expect users to be running that version in production, though
we may not port fixes back before the latest minor version. For example, when
v1.3 comes out, v1.0 will no longer be supported: basically, that means that the
reasonable response to the question "my v1.0 cluster isn't working," is, "you
should probably upgrade it, (and probably should have some time ago)". With
minor releases happening approximately every three months, that means a minor
release is supported for approximately nine months.
This does *not* mean that we expect to introduce breaking changes between v1.0 and v1.3, but it does mean that we probably won't have reasonable confidence in clusters where some components are running at v1.0 and others running at v1.3.
This does *not* mean that we expect to introduce breaking changes between v1.0
and v1.3, but it does mean that we probably won't have reasonable confidence in
clusters where some components are running at v1.0 and others running at v1.3.
This policy is in line with [GKE's supported upgrades policy](https://cloud.google.com/container-engine/docs/clusters/upgrade).
This policy is in line with
[GKE's supported upgrades policy](https://cloud.google.com/container-engine/docs/clusters/upgrade).
## API versioning
@ -91,33 +126,74 @@ This policy is in line with [GKE's supported upgrades policy](https://cloud.goog
Here is an example major release cycle:
* **Kube 1.0 should have API v1 without v1beta\* API versions**
* The last version of Kube before 1.0 (e.g. 0.14 or whatever it is) will have the stable v1 API. This enables you to migrate all your objects off of the beta API versions of the API and allows us to remove those beta API versions in Kube 1.0 with no effect. There will be tooling to help you detect and migrate any v1beta\* data versions or calls to v1 before you do the upgrade.
* The last version of Kube before 1.0 (e.g. 0.14 or whatever it is) will have
the stable v1 API. This enables you to migrate all your objects off of the beta
API versions of the API and allows us to remove those beta API versions in Kube
1.0 with no effect. There will be tooling to help you detect and migrate any
v1beta\* data versions or calls to v1 before you do the upgrade.
* **Kube 1.x may have API v2beta***
* The first incarnation of a new (backwards-incompatible) API in HEAD is v2beta1. By default this will be unregistered in apiserver, so it can change freely. Once it is available by default in apiserver (which may not happen for several minor releases), it cannot change ever again because we serialize objects in versioned form, and we always need to be able to deserialize any objects that are saved in etcd, even between alpha versions. If further changes to v2beta1 need to be made, v2beta2 is created, and so on, in subsequent 1.x versions.
* **Kube 1.y (where y is the last version of the 1.x series) must have final API v2**
* Before Kube 2.0 is cut, API v2 must be released in 1.x. This enables two things: (1) users can upgrade to API v2 when running Kube 1.x and then switch over to Kube 2.x transparently, and (2) in the Kube 2.0 release itself we can cleanup and remove all API v2beta\* versions because no one should have v2beta\* objects left in their database. As mentioned above, tooling will exist to make sure there are no calls or references to a given API version anywhere inside someone's kube installation before someone upgrades.
* Kube 2.0 must include the v1 API, but Kube 3.0 must include the v2 API only. It *may* include the v1 API as well if the burden is not high - this will be determined on a per-major-version basis.
* The first incarnation of a new (backwards-incompatible) API in HEAD is
v2beta1. By default this will be unregistered in apiserver, so it can change
freely. Once it is available by default in apiserver (which may not happen for
several minor releases), it cannot change ever again because we serialize
objects in versioned form, and we always need to be able to deserialize any
objects that are saved in etcd, even between alpha versions. If further changes
to v2beta1 need to be made, v2beta2 is created, and so on, in subsequent 1.x
versions.
* **Kube 1.y (where y is the last version of the 1.x series) must have final
API v2**
* Before Kube 2.0 is cut, API v2 must be released in 1.x. This enables two
things: (1) users can upgrade to API v2 when running Kube 1.x and then switch
over to Kube 2.x transparently, and (2) in the Kube 2.0 release itself we can
cleanup and remove all API v2beta\* versions because no one should have
v2beta\* objects left in their database. As mentioned above, tooling will exist
to make sure there are no calls or references to a given API version anywhere
inside someone's kube installation before someone upgrades.
* Kube 2.0 must include the v1 API, but Kube 3.0 must include the v2 API only.
It *may* include the v1 API as well if the burden is not high - this will be
determined on a per-major-version basis.
#### Rationale for API v2 being complete before v2.0's release
It may seem a bit strange to complete the v2 API before v2.0 is released, but *adding* a v2 API is not a breaking change. *Removing* the v2beta\* APIs *is* a breaking change, which is what necessitates the major version bump. There are other ways to do this, but having the major release be the fresh start of that release's API without the baggage of its beta versions seems most intuitive out of the available options.
It may seem a bit strange to complete the v2 API before v2.0 is released,
but *adding* a v2 API is not a breaking change. *Removing* the v2beta\*
APIs *is* a breaking change, which is what necessitates the major version bump.
There are other ways to do this, but having the major release be the fresh start
of that release's API without the baggage of its beta versions seems most
intuitive out of the available options.
## Patch releases
Patch releases are intended for critical bug fixes to the latest minor version, such as addressing security vulnerabilities, fixes to problems affecting a large number of users, severe problems with no workaround, and blockers for products based on Kubernetes.
Patch releases are intended for critical bug fixes to the latest minor version,
such as addressing security vulnerabilities, fixes to problems affecting a large
number of users, severe problems with no workaround, and blockers for products
based on Kubernetes.
They should not contain miscellaneous feature additions or improvements, and especially no incompatibilities should be introduced between patch versions of the same minor version (or even major version).
They should not contain miscellaneous feature additions or improvements, and
especially no incompatibilities should be introduced between patch versions of
the same minor version (or even major version).
Dependencies, such as Docker or Etcd, should also not be changed unless absolutely necessary, and also just to fix critical bugs (so, at most patch version changes, not new major nor minor versions).
Dependencies, such as Docker or Etcd, should also not be changed unless
absolutely necessary, and also just to fix critical bugs (so, at most patch
version changes, not new major nor minor versions).
## Upgrades
* Users can upgrade from any Kube 1.x release to any other Kube 1.x release as a rolling upgrade across their cluster. (Rolling upgrade means being able to upgrade the master first, then one node at a time. See #4855 for details.)
* However, we do not recommend upgrading more than two minor releases at a time (see [Supported releases](#supported-releases)), and do not recommend running non-latest patch releases of a given minor release.
* Users can upgrade from any Kube 1.x release to any other Kube 1.x release as a
rolling upgrade across their cluster. (Rolling upgrade means being able to
upgrade the master first, then one node at a time. See #4855 for details.)
* However, we do not recommend upgrading more than two minor releases at a
time (see [Supported releases](#supported-releases)), and do not recommend
running non-latest patch releases of a given minor release.
* No hard breaking changes over version boundaries.
* For example, if a user is at Kube 1.x, we may require them to upgrade to Kube 1.x+y before upgrading to Kube 2.x. In others words, an upgrade across major versions (e.g. Kube 1.x to Kube 2.x) should effectively be a no-op and as graceful as an upgrade from Kube 1.x to Kube 1.x+1. But you can require someone to go from 1.x to 1.x+y before they go to 2.x.
* For example, if a user is at Kube 1.x, we may require them to upgrade to
Kube 1.x+y before upgrading to Kube 2.x. In others words, an upgrade across
major versions (e.g. Kube 1.x to Kube 2.x) should effectively be a no-op and as
graceful as an upgrade from Kube 1.x to Kube 1.x+1. But you can require someone
to go from 1.x to 1.x+y before they go to 2.x.
There is a separate question of how to track the capabilities of a kubelet to facilitate rolling upgrades. That is not addressed here.
There is a separate question of how to track the capabilities of a kubelet to
facilitate rolling upgrades. That is not addressed here.
<!-- BEGIN MUNGE: GENERATED_ANALYTICS -->