Automatic merge from submit-queue (batch tested with PRs 47000, 47188, 47094, 47323, 47124)
fix sync loop health check
This PR will do error logging about the fall behind sync for kubelet instead of sync loop healthz checking.
The reason is kubelet can not do sync loop and therefore can not update sync loop time when there is any runtime error, such as docker hung.
When there is any runtime error, according to current implementation, kubelet will not do sync operation and thus kubelet's sync loop time will not be updated. This will make when there is any runtime error, kubelet will also return non 200 response status code when accessing healthz endpoint. This is contrary with #37865 which prevents kubelet from being killed when docker hangs.
**Release note**:
```release-note
fix sync loop health check with seperating runtime errors
```
/cc @yujuhong @Random-Liu @dchen1107
This PR adds the support for allocatable local storage (scratch space).
This feature is only for root file system which is shared by kubernetes
componenets, users' containers and/or images. User could use
--kube-reserved flag to reserve the storage for kube system components.
If the allocatable storage for user's pods is used up, some pods will be
evicted to free the storage resource.
Automatic merge from submit-queue (batch tested with PRs 46635, 45619, 46637, 45059, 46415)
Certificate rotation for kubelet server certs.
Replaces the current kubelet server side self signed certs with certs signed by
the Certificate Request Signing API on the API server. Also renews expiring
kubelet server certs as expiration approaches.
Two Points:
1. With `--feature-gates=RotateKubeletServerCertificate=true` set, the kubelet will
request a certificate during the boot cycle and pause waiting for the request to
be satisfied.
2. In order to have the kubelet's certificate signing request auto approved,
`--insecure-experimental-approve-all-kubelet-csrs-for-group=` must be set on
the cluster controller manager. There is an improved mechanism for auto
approval [proposed](https://github.com/kubernetes/kubernetes/issues/45030).
**Release note**:
```release-note
With `--feature-gates=RotateKubeletServerCertificate=true` set, the kubelet will
request a server certificate from the API server during the boot cycle and pause
waiting for the request to be satisfied. It will continually refresh the certificate as
the certificates expiration approaches.
```
Replaces the current kubelet server side self signed certs with certs
signed by the Certificate Request Signing API on the API server. Also
renews expiring kubelet server certs as expiration approaches.
Automatic merge from submit-queue (batch tested with PRs 46124, 46434, 46089, 45589, 46045)
Support TCP type runtime endpoint for kubelet
**What this PR does / why we need it**:
Currently the grpc server for kubelet and dockershim has a hardcoded endpoint: unix socket '/var/run/dockershim.sock', which is not applicable on non-unix OS.
This PR is to support TCP endpoint type besides unix socket.
**Which issue this PR fixes**
This is a first attempt to address issue https://github.com/kubernetes/kubernetes/issues/45927
**Special notes for your reviewer**:
Before this change, running on Windows node results in:
```
Container Manager is unsupported in this build
```
After adding the cm stub, error becomes:
```
listen unix /var/run/dockershim.sock: socket: An address incompatible with the requested protocol was used.
```
This PR is to fix those two issues.
After this change, still meets 'seccomp' related issue when running on Windows node, needs more updates later.
**Release note**:
Automatic merge from submit-queue
Fix some typo of comment in kubelet.go
**What this PR does / why we need it**:
The PR is to fix some typo in kubelet.go
**Which issue this PR fixes** *(optional, in `fixes #<issue number>(, fixes #<issue_number>, ...)` format, will close that issue when PR gets merged)*: fixes #
N/A
**Special notes for your reviewer**:
**Release note**:
```release-note
```
Automatic merge from submit-queue (batch tested with PRs 46022, 46055, 45308, 46209, 43590)
Eviction does not evict unless the previous pod has been cleaned up
Addresses #43166
This PR makes two main changes:
First, it makes the eviction loop re-trigger immediately if there may still be pressure. This way, if we already waited 10 seconds to delete a pod, we dont need to wait another 10 seconds for the next synchronize call.
Second, it waits for the pod to be cleaned up (including volumes, cgroups, etc), before moving on to the next synchronize call. It has a timeout for this operation currently set to 30 seconds.
Automatic merge from submit-queue
Reorganize kubelet tree so apis can be independently versioned
@yujuhong @lavalamp @thockin @bgrant0607
This is an example of how we might reorganize `pkg/kubelet` so the apis it exposes can be independently versioned. This would also provide a logical place to put the `KubeletConfiguration` type, which currently lives in `pkg/apis/componentconfig`; it could live in e.g. `pkg/kubelet/apis/config` instead.
Take a look when you have a chance and let me know what you think. The most significant change in this PR is reorganizing `pkg/kubelet/api` to `pkg/kubelet/apis`, the rest is pretty much updating import paths and `BUILD` files.
Automatic merge from submit-queue (batch tested with PRs 45685, 45572, 45624, 45723, 45733)
Remove unused fields from Kubelet struct
Just a small attempt to clean up some unused fields in the kubelet struct. This doesn't make any actual code changes.
/assign @mtaufen
Automatic merge from submit-queue
Enable shared PID namespace by default for docker pods
**What this PR does / why we need it**: This PR enables PID namespace sharing for docker pods by default, bringing the behavior of docker in line with the other CRI runtimes when used with docker >= 1.13.1.
**Which issue this PR fixes** *(optional, in `fixes #<issue number>(, fixes #<issue_number>, ...)` format, will close that issue when PR gets merged)*: ref #1615
**Special notes for your reviewer**: cc @dchen1107 @yujuhong
**Release note**:
```release-note
Kubernetes now shares a single PID namespace among all containers in a pod when running with docker >= 1.13.1. This means processes can now signal processes in other containers in a pod, but it also means that the `kubectl exec {pod} kill 1` pattern will cause the pod to be restarted rather than a single container.
```
Automatic merge from submit-queue (batch tested with PRs 45453, 45307, 44987)
Migrate the docker client code from dockertools to dockershim
Move docker client code from dockertools to dockershim/libdocker. This includes
DockerInterface (renamed to Interface), FakeDockerClient, etc.
This is part of #43234
Automatic merge from submit-queue (batch tested with PRs 45508, 44258, 44126, 45441, 45320)
cloud initialize node in external cloud controller
@thockin This PR adds support in the `cloud-controller-manager` to initialize nodes (instead of kubelet, which did it previously)
This also adds support in the kubelet to skip node cloud initialization when `--cloud-provider=external`
Specifically,
Kubelet
1. The kubelet has a new flag called `--provider-id` which uniquely identifies a node in an external DB
2. The kubelet sets a node taint - called "ExternalCloudProvider=true:NoSchedule" if cloudprovider == "external"
Cloud-Controller-Manager
1. The cloud-controller-manager listens on "AddNode" events, and then processes nodes that starts with that above taint. It performs the cloud node initialization steps that were previously being done by the kubelet.
2. On addition of node, it figures out the zone, region, instance-type, removes the above taint and updates the node.
3. Then periodically queries the cloudprovider for node addresses (which was previously done by the kubelet) and updates the node if there are new addresses
```release-note
NONE
```
Automatic merge from submit-queue
adds log when gpuManager.start() failed
If gpuManager.start() returns error, there is no log.
We confused with scheduler do not schedule any pod(with gpu) to one node.
kubectl describe node xxx shows there is no gpu on that node, because the gpu driver do not work on that node, gpuManager.start() failed, but we can not see anything in log.
Automatic merge from submit-queue (batch tested with PRs 45316, 45341)
Pass NoOpLegacyHost to dockershim in --experimental-dockershim mode
This allows dockershim to use network plugins, if needed.
/cc @Random-Liu
This commit deletes code in dockertools that is only used by
DockerManager. A follow-up change will rename and clean up the rest of
the files in this package.
The commit also sets EnableCRI to true if the container runtime is not
rkt. A follow-up change will remove the flag/field and all references to
it.
rktnetes is not a CRI implementation, and does not provide runtime
conditions. This change fixes the issue where rkt will never be
considered running from kubelet's point of view.
Make the location of dockershim.sock configurable, so downstream
projects (such as OpenShift) can place it in a location that does not
require root access (e.g. for integration tests).
Make the kubelet respect and use the values of
--container-runtime-endpoint and --image-service-endpoint, if set. If
unset, the default value of /var/run/dockershim.sock is used.
Automatic merge from submit-queue
Clearer ImageGC failure errors. Fewer events.
Addresses #26000. Kubelet often "fails" image garbage collection if cAdvisor has not completed the first round of stats collection. Don't create events for a single failure, and make log messages more specific.
@kubernetes/sig-node-bugs
Kubelet flags are not necessarily appropriate for the KubeletConfiguration
object. For example, this PR also removes HostnameOverride and NodeIP
from KubeletConfiguration. This is a preleminary step to enabling Nodes
to share configurations, as part of the dynamic Kubelet configuration
feature (#29459). Fields that must be unique for each node inhibit
sharing, because their values, by definition, cannot be shared.
Automatic merge from submit-queue
[Bug] Fix gpu initialization in Kubelet
Kubelet incorrectly fails if `AllAlpha=true` feature gate is enabled with container runtimes that are not `docker`.
Replaces #42407
Automatic merge from submit-queue (batch tested with PRs 41931, 39821, 41841, 42197, 42195)
Use `docker logs` directly if the docker logging driver is not `json-file`
Fixes https://github.com/kubernetes/kubernetes/issues/41996.
Post the PR first, I still need to manually test this, because we don't have test coverage for journald logging pluggin.
@yujuhong @dchen1107
/cc @kubernetes/sig-node-pr-reviews
Introduced chages:
1. Re-writing of the resolv.conf file generated by docker.
Cluster dns settings aren't passed anymore to docker api in all cases, not only for pods with host network:
the resolver conf will be overwritten after infra-container creation to override docker's behaviour.
2. Added new one dnsPolicy - 'ClusterFirstWithHostNet', so now there are:
- ClusterFirstWithHostNet - use dns settings in all cases, i.e. with hostNet=true as well
- ClusterFirst - use dns settings unless hostNetwork is true
- Default
Fixes#17406
Automatic merge from submit-queue
Extend experimental support to multiple Nvidia GPUs
Extended from #28216
```release-note
`--experimental-nvidia-gpus` flag is **replaced** by `Accelerators` alpha feature gate along with support for multiple Nvidia GPUs.
To use GPUs, pass `Accelerators=true` as part of `--feature-gates` flag.
Works only with Docker runtime.
```
1. Automated testing for this PR is not possible since creation of clusters with GPUs isn't supported yet in GCP.
1. To test this PR locally, use the node e2e.
```shell
TEST_ARGS='--feature-gates=DynamicKubeletConfig=true' FOCUS=GPU SKIP="" make test-e2e-node
```
TODO:
- [x] Run manual tests
- [x] Add node e2e
- [x] Add unit tests for GPU manager (< 100% coverage)
- [ ] Add unit tests in kubelet package
Automatic merge from submit-queue (batch tested with PRs 41994, 41969, 41997, 40952, 40576)
Guaranteed admission for Critical Pods
This is the first step in implementing node-level preemption for critical pods.
It defines the AdmissionFailureHandler interface, which allows callers, like the kubelet, to define how failed predicates are handled, and take steps to correct failures if necessary.
In the kubelet's implementation, it triggers preemption if the pod being admitted is critical, and if the only failed predicates are InsufficientResourceErrors, then it prempts (not yet implemented) other other pods to allow admission of the critical pod.
cc: @vishh
Where possible, switch the scheduler to use generated listers and
informers. There are still some places where it probably makes more
sense to use one-off reflectors/informers (listing/watching just a
single node, listing/watching scheduled & unscheduled pods using a field
selector).
Automatic merge from submit-queue (batch tested with PRs 38101, 41431, 39606, 41569, 41509)
Report node not ready on failed PLEG health check
Report node not ready if PLEG health check fails.
Automatic merge from submit-queue (batch tested with PRs 38101, 41431, 39606, 41569, 41509)
optimize killPod() and syncPod() functions
make sure that one of the two arguments must be non-nil: runningPod, status ,just like the function note says
and judge the return value in syncPod() function before setting podKilled
Automatic merge from submit-queue (batch tested with PRs 41466, 41456, 41550, 41238, 41416)
Delay Deletion of a Pod until volumes are cleaned up
#41436 fixed the bug that caused #41095 and #40239 to have to be reverted. Now that the bug is fixed, this shouldn't cause problems.
@vishh @derekwaynecarr @sjenning @jingxu97 @kubernetes/sig-storage-misc
make sure that one of the two arguments must be non-nil: runningPod, status ,just like the function note says
and judge the return value in syncPod() function before setting podKilled
Automatic merge from submit-queue
Allow multipe DNS servers as comma-seperated argument for kubelet --dns
This PR explores how kubectls "--dns" could be extended to specify multiple DNS servers for in-cluster PODs. Testing on the local libvirt-coreos cluster shows that multiple DNS server are injected without issues.
Specifying multiple DNS servers increases resilience against
- Packet drops
- Single server failure
I am debugging services that do 50+ DNS requests for a single incoming interactive request, thus highly increase the chance of a slowdown (+5s) due to a single packet drop. Switching to two DNS servers will reduce the impact of the issues (roughly +1s on glibc, 0s on musl, error-rate goes down to error-rate^2).
Note that there is no need to change any runtime related code as far as I know. In the case of "default" dns the /etc/resolv.conf is parsed and multiple DNS server are send to the backend anyway. This only adds the same capability for the clusterFirst case.
I've heard from @thockin that multiple DNS entries are somehow considered. I've no idea what was considered, though. This is what I would like to see for our production use, though.
```release-note
NONE
```
This change makes kubelet to use the CRI implementation by default,
unless the users opt out explicitly by using --enable-cri=false.
For the rkt integration, the --enable-cri flag will have no effect
since rktnetes does not use CRI.
Also, mark the original --experimental-cri flag hidden and deprecated,
so that we can remove it in the next release.
Automatic merge from submit-queue (batch tested with PRs 40796, 40878, 36033, 40838, 41210)
Implement TTL controller and use the ttl annotation attached to node in secret manager
For every secret attached to a pod as volume, Kubelet is trying to refresh it every sync period. Currently Kubelet has a ttl-cache of secrets of its pods and the ttl is set to 1 minute. That means that in large clusters we are targetting (5k nodes, 30pods/node), given that each pod has a secret associated with ServiceAccount from its namespaces, and with large enough number of namespaces (where on each node (almost) every pod is from a different namespace), that resource in ~30 GETs to refresh all secrets every minute from one node, which gives ~2500QPS for GET secrets to apiserver.
Apiserver cannot keep up with it very easily.
Desired solution would be to watch for secret changes, but because of security we don't want a node watching for all secrets, and it is not possible for now to watch only for secrets attached to pods from my node.
So as a temporary solution, we are introducing an annotation that would be a suggestion for kubelet for the TTL of secrets in the cache and a very simple controller that would be setting this annotation based on the cluster size (the large cluster is, the bigger ttl is).
That workaround mean that only very local changes are needed in Kubelet, we are creating a well separated very simple controller, and once watching "my secrets" will be possible it will be easy to remove it and switch to that. And it will allow us to reach scalability goals.
@dchen1107 @thockin @liggitt
Automatic merge from submit-queue (batch tested with PRs 40873, 40948, 39580, 41065, 40815)
[CRI] Enable Hostport Feature for Dockershim
Commits:
1. Refactor common hostport util logics and add more tests
2. Add HostportManager which can ADD/DEL hostports instead of a complete sync.
3. Add Interface for retreiving portMappings information of a pod in Network Host interface.
Implement GetPodPortMappings interface in dockerService.
4. Teach kubenet to use HostportManager
Automatic merge from submit-queue (batch tested with PRs 41103, 41042, 41097, 40946, 40770)
Use Clientset interface in KubeletDeps
**What this PR does / why we need it**:
This replaces the Clientset struct with the equivalent interface for the KubeClient injected via KubeletDeps. This is useful for testing and for accessing the Node and Pod status event stream without an API server.
**Special notes for your reviewer**:
Follow up to #4907
**Release note**:
`NONE`
Automatic merge from submit-queue (batch tested with PRs 40289, 40877, 40879, 39972, 40942)
Rename experimental-cgroups-per-pod flag
**What this PR does / why we need it**:
1. Rename `experimental-cgroups-per-qos` to `cgroups-per-qos`
1. Update hack/local-up-cluster to match `CGROUP_DRIVER` with docker runtime if used.
**Special notes for your reviewer**:
We plan to roll this feature out in the upcoming release. Previous node e2e runs were running with this feature on by default. We will default this feature on for all e2es next week.
**Release note**:
```release-note
Rename --experiemental-cgroups-per-qos to --cgroups-per-qos
```
Automatic merge from submit-queue
Optionally avoid evicting critical pods in kubelet
For #40573
```release-note
When feature gate "ExperimentalCriticalPodAnnotation" is set, Kubelet will avoid evicting pods in "kube-system" namespace that contains a special annotation - `scheduler.alpha.kubernetes.io/critical-pod`
This feature should be used in conjunction with the rescheduler to guarantee availability for critical system pods - https://kubernetes.io/docs/admin/rescheduler/
```
Depending on an exact cluster setup multiple dns may make sense.
Comma-seperated lists of DNS server are quite common as DNS servers
are always plain IPs.
- split out port forwarding into its own package
Allow multiple port forwarding ports
- Make it easy to determine which port is tied to which channel
- odd channels are for data
- even channels are for errors
- allow comma separated ports to specify multiple ports
Add portfowardtester 1.2 to whitelist
Automatic merge from submit-queue (batch tested with PRs 40232, 40235, 40237, 40240)
move listers out of cache to reduce import tree
Moving the listers from `pkg/client/cache` snips links to all the different API groups from `pkg/storage`, but the dreaded `ListOptions` remains.
@sttts
Automatic merge from submit-queue (batch tested with PRs 37228, 40146, 40075, 38789, 40189)
kubelet: storage: teardown terminated pod volumes
This is a continuation of the work done in https://github.com/kubernetes/kubernetes/pull/36779
There really is no reason to keep volumes for terminated pods attached on the node. This PR extends the removal of volumes on the node from memory-backed (the current policy) to all volumes.
@pmorie raised a concern an impact debugging volume related issues if terminated pod volumes are removed. To address this issue, the PR adds a `--keep-terminated-pod-volumes` flag the kubelet and sets it for `hack/local-up-cluster.sh`.
For consideration in 1.6.
Fixes#35406
@derekwaynecarr @vishh @dashpole
```release-note
kubelet tears down pod volumes on pod termination rather than pod deletion
```
Automatic merge from submit-queue (batch tested with PRs 40168, 40165, 39158, 39966, 40190)
dockershim: add support for the 'nsenter' exec handler
This change simply plumbs the kubelet configuration
(--docker-exec-handler) to DockerService.
This fixes#35747.
- exclude duplicates while merging of host's and dns' search lines to form pod's one
- truncate pod's search line if it exceeds resolver limits: is > 255 chars and containes > 6 searches
- monitoring the resolv.conf file which is used by kubelet (set thru --resolv-conf="") and logging and eventing if search line in it consists of more than 3 entries
(or 6 if Cluster Domain is set) or its lenght is > 255 chars
- logging and eventing when a pod's search line is > 255 chars or containes > 6 searches during forming
Fixes#29270
Automatic merge from submit-queue
kubelet: remove the pleg health check from healthz
This prevents kubelet from being killed when docker hangs.
Also, kubelet will report node not ready if PLEG hangs (`docker ps` + `docker inspect`).
Automatic merge from submit-queue (batch tested with PRs 39079, 38991, 38673)
Support systemd based pod qos in CRI dockershim
This PR makes pod level QoS works for CRI dockershim for systemd based cgroups. And will also fix#36807
- [x] Add cgroupDriver to dockerService and use docker info api to set value for it
- [x] Add a NOTE that detection only works for docker 1.11+, see [CHANGE LOG](https://github.com/docker/docker/blob/master/CHANGELOG.md#1110-2016-04-13)
- [x] Generate cgroupParent in syntax expected by cgroupDriver
- [x] Set cgroupParent to hostConfig for both sandbox and user container
- [x] Check if kubelet conflicts with cgroup driver of docker
cc @derekwaynecarr @vishh
Automatic merge from submit-queue (batch tested with PRs 38154, 38502)
Rename "release_1_5" clientset to just "clientset"
We used to keep multiple releases in the main repo. Now that [client-go](https://github.com/kubernetes/client-go) does the versioning, there is no need to keep releases in the main repo. This PR renames the "release_1_5" clientset to just "clientset", clientset development will be done in this directory.
@kubernetes/sig-api-machinery @deads2k
```release-note
The main repository does not keep multiple releases of clientsets anymore. Please find previous releases at https://github.com/kubernetes/client-go
```
Automatic merge from submit-queue (batch tested with PRs 37208, 37446, 37420)
Kubelet log modification
Keep in line with the other error logs in the function.
After return, the caller records the error log.Delete redundant logs
Automatic merge from submit-queue
Function annotation modification
“return kl.pleg.Healthy()”，Based on the return function,"healty" to "healthy" better
Automatic merge from submit-queue
kubelet: don't reject pods without adding them to the pod manager
kubelet relies on the pod manager as a cache of the pods in the apiserver (and
other sources) . The cache should be kept up-to-date even when rejecting pods.
Without this, kubelet may decide at any point to drop the status update
(request to the apiserver) for the rejected pod since it would think the pod no
longer exists in the apiserver.
This should fix#37658
Automatic merge from submit-queue
remove checking mount point in cleanupOrphanedPodDirs
To avoid nfs hung problem, remove the mountpoint checking code in
cleanupOrphanedPodDirs(). This removal should still be safe because it checks whether there are still directories under pod's volume and if so, do not delete the pod directory.
Note: After removing the mountpoint check code in cleanupOrphanedPodDirs(), the directories might not be cleaned up in such situation.
1. delete pod, kubelet reconciler tries to unmount the volume directory successfully
2. before reconciler tries to delete the volume directory, kubelet gets retarted
3. since under pod directory, there are still volume directors exist (but not mounted), cleanupOrphanedPodDIrs() will not clean them up.
Will work on a follow up PR to solve above issue.
