kubernetes-sigs / lws / 36310016928

88%

Build Type

push

github

Committed by web-flow

Commit Message

fix: reconcile LeaderWorkerSet on member pod deletion to clean up obsolete PodGroups (#1096)

When using the kubernetes scheduler provider, scaling down or scaling to
zero could leave obsolete PodGroup objects in the cluster whenever a worker
pod was the last member pod to terminate. This occurred because
LeaderWorkerSetReconciler.SetupWithManager previously filtered out all
worker pod events inside its mapping function, dropping the final DeleteEvent.

This change:
1. Replaces the inline map function with enqueueLWSRequests so that member pod
   deletions can trigger reconciliation for the LeaderWorkerSet.
2. Extracts podWatchPredicate into a separate method with explicit predicates:
   - CreateFunc: only leader pods (worker-index: "0") trigger reconciliation.
   - UpdateFunc: only leader pods with changes to
     GroupRestartBudgetExhaustedAnnotationKey trigger reconciliation.
   - DeleteFunc: leader pods always trigger reconciliation; worker pods only
     trigger reconciliation if KubernetesProvider is active and the pod has
     spec.schedulingGroup or WorkloadSchedulingAnnotationKey.
   - GenericFunc: explicitly set to return false.
3. Adds unit tests for podWatchPredicate and extends TestEnqueueLWSRequests.
4. Adds an integration test in workload_scheduling_test.go verifying PodGroup
   cleanup when a worker pod is the last to be deleted during scale-down.

Signed-off-by: Harry Li <harrymingh@gmail.com>

Coverage Stats

30 of 34 new or added lines in 1 file covered. (88.24%)

4 existing lines in 1 file now uncovered.

5805 of 6883 relevant lines covered (84.34%)

37.06 hits per line