github/kubernetes
1
0
Fork 0
mirror of https://github.com/k3s-io/kubernetes.git synced 2026-07-17 20:00:07 +00:00
Code Issues Projects Releases Wiki Activity

Merge pull request #135489 from ania-borowiec/update_comment

Browse Source 
Update async preemption comment to reflect the current state of the code
This commit is contained in:
Kubernetes Prow Robot
2025-12-17 22:36:13 -08:00
committed by GitHub
parent c5a0c31294 0cf3d0e20a
commit 62db4db266
1 changed files with 8 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
pkg/scheduler/framework/preemption/preemption.go
View File

@@ -593,15 +593,16 @@ func (ev *Evaluator) prepareCandidateAsync(c Candidate, pod *v1.Pod, pluginName
}
if len(victimPods) > 1 {
// We can evict all victims in parallel, but the last one.
// We have to remove the pod from the preempting map before the last one is evicted
// because, otherwise, the pod removal might be notified to the scheduling queue before
// we remove this pod from the preempting map,
// and the pod could end up stucking at the unschedulable pod pool
// by all the pod removal events being ignored.
// In order to prevent requesting preemption of the same pod multiple times for the same preemptor,
// preemptor is marked as "waiting for preemption of a victim" (by adding it to preempting map).
// For optimization purposes the last victim is recorded in lastVictimsPendingPreemption.
// We can evict all victims in parallel, but the last one.
// While deleting the last victim, we want the PreEnqueue check to be able to verify if the eviction
// is in fact ongoing, or if it has already completed, but the function has not returned yet.
// In the latter case, PreEnqueue (in `IsPodRunningPreemption`) reads the state of the last victim in
// `lastVictimsPendingPreemption` and does not block the pod.
// This helps us avoid the situation where pod removal might be notified to the scheduling queue before
// the preemptor completes the deletion API calls and is removed from the `preempting` map - that way
// the preemptor could end up stuck in the unschedulable pool, with all pod removal events being ignored.
ev.Handler.Parallelizer().Until(ctx, len(victimPods)-1, preemptPod, ev.PluginName)
if err := errCh.ReceiveError(); err != nil {
utilruntime.HandleErrorWithContext(ctx, err, "Error occurred during async preemption")