kubernetes-sigs / karpenter / 30001588366

82%

main: 82%

Build Type

Pull #3168

github

Committed by maks3201

Commit Message

fix: respect pod terminationGracePeriodSeconds when waiting after drain

MinDrainTime (5s) is used as the minimum wait between all pods being
evicted (deletionTimestamp set) and calling cloudProvider.Delete().
However, IsWaitingEviction() is a pure etcd/API check - it returns
false as soon as eviction is accepted, while the kubelet may still be
running the preStop hook for the pod's full terminationGracePeriodSeconds.

With the default tgps of 30s, Karpenter was calling EC2 TerminateInstances
~25s before preStop hooks could complete, causing:
  FailedPreStopHook
  FailedKillPod: ttrpc: closed

Fix: compute effectiveDrainTime as max(MinDrainTime, maxTGPS) across
pods still present on the node with deletionTimestamp set. MinDrainTime
is retained as the floor to preserve the original behavior of catching
late-scheduled pods.

Fixes: #1894

Pull Request Pull Request #3168: fix: respect pod terminationGracePeriodSeconds when waiting after drain

Coverage Stats

14 of 16 new or added lines in 1 file covered. (87.5%)

37 existing lines in 1 file now uncovered.

16582 of 20112 relevant lines covered (82.45%)

97538.1 hits per line