When watch.Broadcaster.Shutdown() is called it drains all queued events
then calls closeAll(), which closes every watcher's result channel.
eventBroadcasterImpl.Shutdown() calls Broadcaster.Shutdown() first,
then calls the cancellation context's cancel() function. Between those
two steps there is a window in which the result channel is closed while
the cancellation context is still live.
Without the two-value channel receive the goroutine in StartEventWatcher
would spin on the already-closed channel: each select iteration
immediately receives the zero-value watch.Event, the type assertion
fails (nil interface, ok == false), and the loop continues burning CPU
until the select scheduler eventually picks the cancelationCtx.Done()
case.
Guard against this by reading the ok boolean from the channel receive:
case watchEvent, ok := <-watcher.ResultChan():
if !ok {
return
}
This is the correct and idiomatic Go pattern for a channel that may be
closed by its producer. Note that when this return path is taken the
broadcaster has already delivered every queued event (Broadcaster.Shutdown
blocks until the distribute loop exits before closeAll runs), so no
events are silently dropped.
Add a regression test (TestStartEventWatcherExitsOnDirectShutdown) that
creates a broadcaster without an external context so Shutdown() is
fully synchronous, starts a watcher, and verifies the goroutine exits
cleanly via goleak.VerifyNone.
Signed-off-by: Rajneesh180 <rajneeshrehsaan48@gmail.com>
Kubernetes-commit: 95c15b54069922b0a66c198a064577ea0a160694
[Declarative Validation] Bring `k8s:maxLength` tag in line with OpenAPI `maxLength` validation semantics
Kubernetes-commit: e08e598df07bc929679ef046418992a8205da18f
* Promote MutableCSINodeAllocatableCount to GA
Signed-off-by: Eddie Torres <torredil@amazon.com>
* Lock MutableCSINodeAllocatableCount feature gate to default
Signed-off-by: Eddie Torres <torredil@amazon.com>
---------
Signed-off-by: Eddie Torres <torredil@amazon.com>
Kubernetes-commit: 41bb4b6a8b4e0f15ab7ffcc6369c68bc599fb957
The change introduced in 59cd1d0b3bb378f40a639e21b615f4df1d4a5a14
causes PollUntilContextTimeout to call the condition function even when
the context is canceled already. This is unnecessary in our case and
only leads to one extra loop of processing and extra error messages.
This change makes the renew loop behave like it did before the change.
Kubernetes-commit: aa494f8174d745c3d9842f3d48272a3c2d6dcbcc
In practice, TimeAdded is managed by the API server. When admins used
DeviceTaintRule to simulate eviction, then change the effect to really evict,
it is useful to calculate tolerations based on the time when that second
update happened. Therefore the TimeAdded field gets bumped automatically
when changing the effect.
Kubernetes-commit: f28dc4139208e64fe2882cd44d548c25020f2e3a
The error returned from Until() is solely from context cancellation
which is expected behavior when the reflector is stopped. Logging
this as an error (or even at V(4)) creates unnecessary noise.
Kubernetes-commit: cc483208aa306b8c4078d4118cf78a10e58481ec
It's GA now. To regenerate the files I did:
make WHAT=cmd/kube-apiserver
make update
While we are there, remove the reference that the field is alpha.
Signed-off-by: Rodrigo Campos <rodrigo@amutable.com>
Kubernetes-commit: 8db51091e82490bccc0018763ba7e05ffb8ab458
This is just a sed to remove the old data from the text fixtures too.
While two files are clearly test data given that they include testdata
in their name, these two is not so obvious but are also test data:
* openapi/swagger-with-shared-parameters.json: is referenced in "staging/src/k8s.io/cli-runtime/pkg/resource/query_param_verifier_test.go",
in particular in this part: ´filepath.Join("..", "..", "artifacts", "openapi", "swagger-with-shared-parameters.json")´
* batch.k8s.io_v1.json: is in
`staging/src/k8s.io/kubectl/pkg/explain/v2/templates/plaintext_test.go`,
in a "//go:embed batch.k8s.io_v1.json".
Signed-off-by: Rodrigo Campos <rodrigo@amutable.com>
Kubernetes-commit: 8d0f80e4ed87658c0e05b9db0690927547be6ad5
