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

more

Browse Source
This commit is contained in:
Chao Xu 2016-09-20 11:22:05 -07:00
parent d67eb53db2
commit b58746f603
1 changed files with 29 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

37
docs/proposals/synchronous-garbage-collection.md
View File

@ -1,6 +1,6 @@
# Overview
Some users of the server-side garbage collection need to tell if the garbage collection is done. Synchronous Garbage Collection is a best-effort (see [unhandled cases](#unhandled-cases)) mechanism to enable such use cases: after the API server receives a deletion request of an owning object, the object keeps existing in the key-value store until all its dependents are deleted by the garbage collector.
Some users of the server-side garbage collection need to tell if the garbage collection is done ([example](https://github.com/kubernetes/kubernetes/issues/19701#issuecomment-236997077)). Synchronous Garbage Collection is a best-effort (see [unhandled cases](#unhandled-cases)) mechanism to enable such use cases: after the API server receives a deletion request of an owning object, the object keeps existing in the key-value store until all its dependents are deleted from the key-value store by the garbage collector.
Tracking issue: https://github.com/kubernetes/kubernetes/issues/29891
@ -13,7 +13,7 @@ We need to make changes in the API, the API Server, and the garbage collector to
DeleteOptions {
…
// If SynchronousGarbageCollection is set, the object will not be deleted immediately. Instead, a GarbageCollectionInProgress finalizer will be placed on the object. The garbage collector will remove the finalizer from the object when all depdendents are deleted.
// SynchronousGarbageCollection is ignored if OrphanDependents is true or nil.
// SynchronousGarbageCollection and OrphanDependents are exclusive.
// SynchronousGarbageCollection default to false.
// SynchronousGarbageCollection is cascading, i.e., the object’s dependents will be deleted with the same SynchronousGarbageCollection.
SynchronousGarbageCollection *bool
@ -46,7 +46,7 @@ Delete() function needs to check the DeleteOptions.SynchronousGarbageCollection.
* Ignores the object if it doesn’t exist in `uidToNode`, or if the object is not marked as “GC in progress” in `uidToNode`.
* To avoid racing with another controller, it requeues the object if `observedGeneration < Generation`. This is best-effort, see [unhandled cases](#unhandled-cases).
* Checks if the object has dependents
* If not, send a PUT request to remove the GCFinalizer
* If not, send a PUT request to remove the `GCFinalizer`
* If so, then add all dependents to the `dirtryQueue`; we need bookkeeping to avoid adding the dependents repeatedly if the owner gets in the `synchronousGC queue` multiple times.
**Modifications to processItem()**
@ -55,19 +55,40 @@ Delete() function needs to check the DeleteOptions.SynchronousGarbageCollection.
* treat an owner as "not exist" if `owner.DeletionTimestamp != nil && !owner.Finalizers.Has(OrphanFinalizer)`, otherwise Synchronous GC will not progress because the owner keeps existing in the key-value store.
* when deleting dependents, it should use the same `DeleteOptions.SynchronousGC` as the owner’s finalizers suggest.
* if an object has multiple owners, some owners still exit while other owners are in the synchronous GC stage, then according to the existing logic of GC, the object wouldn't be deleted. To unblock the synchronous GC of those owners, `processItem()` has to remove the ownerReferences pointing to them.
**Handling circular dependencies**
SynchronousGC will enter a deadlock in the presence of circular dependencies. To break the deadlock, we need to timeout a GCFinalizer. To implement the timeout, GC adds an object that has a GCFinalizer into a [delaying queue](../../pkg/util/workqueue/delaying_queue.go) when it's observed, and removes the GCFinalizer from it when the time is up. The timeout value should be proportional to the number of dependents, including indirect ones.
SynchronousGC will enter a deadlock in the presence of circular dependencies. Here are two alternative approaches to break the deadlock:
1. Timeout a `GCFinalizer `: To implement the timeout, GC adds an object that has a `GCFinalizer` into a [delaying queue](../../pkg/util/workqueue/delaying_queue.go) when it's observed, and removes the `GCFinalizer` from it when the time is up. The timeout value should be proportional to the number of dependents, including indirect ones.
2. Lazily detecting circular dependencies: when `processItem()` processes an object, if it finds the object and all of its owners have the `GCFinalizer`, it searches the internal owner-dependency relationship graph (`uidToNode`) to check if the object and any of its owner are in a circle where all objects have the `GCFinalizer`. If so, it removes the `GCFinzlier` from the object to break the circle.
## Unhandled cases
* If the GC observes the owning object with the GCFinalizer before it observes the creation of all the dependents, GC will remove the finalizer from the owning object before all dependents are gone. Hence, “Synchronous GC” is best-effort, though we guarantee that the dependents will be deleted eventually. We face a similar case when handling OrphanFinalizer, see [GC known issues](https://github.com/kubernetes/kubernetes/issues/26120).
* If the GC observes the owning object with the `GCFinalizer` before it observes the creation of all the dependents, GC will remove the finalizer from the owning object before all dependents are gone. Hence, “Synchronous GC” is best-effort, though we guarantee that the dependents will be deleted eventually. We face a similar case when handling OrphanFinalizer, see [GC known issues](https://github.com/kubernetes/kubernetes/issues/26120).
## Implications to existing clients
**Namespce controller** can [handle finalizers](https://github.com/kubernetes/kubernetes/pull/32524), so it can properly delete a namespace if there is synchronous GC going on in the namespace. Also, we can convert namespace controller to use synchronous GC.
Finalizer breaks an assumption that many Kubernetes components have: a deletion request with `grace period=0` will immediately remove the object from the key-value store. This is not true if an object has pending finalizers, the object will continue to exist, and currently the API server will not return an error in this case.
We should be able to convert **kubectl delete** reapers to use synchronous GC.
**Namespce controller** suffered from this [problem](https://github.com/kubernetes/kubernetes/issues/32519) and was fixed in [#32524](https://github.com/kubernetes/kubernetes/pull/32524) by retrying every 15s if there are objects with pending finalizers to be removed from the key-value store. Object with pending `GCFinalizer` might take arbitrary long time be deleted, so namespace deletion might time out.
For other clients, they are able to work with synchronous GC as long as they can cope with finalizers in general.
**kubelet** deletes the pod from the key-value store after all its containers are terminated ([code](https://github.com/kubernetes/kubernetes/blob/master/pkg/kubelet/status/status_manager.go#L441-L443)). It also assumes if API server does not return an error, the pod is removed from the key-value store. Breaking the assumption will not break `kubelet` though, because the `pod` must have already been in the terminated `phase`, `kubelet` will not care to manage it.
**Node controller** forcefully deletes pod if the pod is scheduled to a node that does not exist ([code](https://github.com/kubernetes/kubernetes/blob/master/pkg/controller/node/nodecontroller.go#L474)). The pod will continue to exist if it has pending finalizers. The node controller will futilely retry the deletion. The `node controller` forcefully deletes pods before deleting the node ([code](https://github.com/kubernetes/kubernetes/blob/master/pkg/controller/node/nodecontroller.go#L592)). If the pods have pending finalizers, the `node controller` will go ahead deleting the node, leaving those pods behind. Other components will take care of the pending finalizers.
**Podgc** deletes terminated pods if there are too many of them in the cluster. `Podgc` should remove any pending finalizers to make sure the pods are deleted.
**Deployment controller** adopts existing `ReplicaSet` (RS) if its template matches. If a matching RS has a pending `GCFinalizer`, deployment shouldn't adopt it, because the RS controller will scale up/down a RS that's being deleted. Hence, `deployment controller` needs to check if a RS is being deleted before adopting it. If the RS is being deleted, then the `deployment controller` should wait for the status of the RS showing 0 replicas, and then create a new RS.
**Replication controller manager**, **Job controller**, and **ReplicaSet controller** send deletion request of pods. `kubelet` will drive these pods to the terminated phase, so the pods will be ignored by the controllers even if they keep existing in the key-value store because of pending finalizers.
**Endpoints controller** deletes endpoints. It does not double check if the endpoint is gone, so
One usage of the synchronous GC is to replace the **kubectl delete** reapers. Currently `kubectl delete` blocks until all dependents and the owner are deleted. To maintain this behavior, after switched to using synchronous GC, *kubectl delete* needs to poll on the removal of the owner object.
## Security implications
A user who are authorized to update one object can affect the synchronous GC behavior of another object. Specifically, even if a user is only authorized to update a pod, he can set another object as the pod's owner, and set a very long grace termination period for the pod, then he makes the synchronous GC of the owner takes long time.