kubernetes-sigs / karpenter / 26238594231

81%

main: 81%

Build Type

Pull #3038

github

Committed by praveen9354

Commit Message

fix(termination): prevent negative pod grace period when Node Repair triggers

Node Repair (node health controller) annotated NodeClaims with
NodeClaimTerminationTimestampAnnotationKey = now(). By the time the
termination reconciler ran, the timestamp was already in the past, so
time.Until(annotationTime) was negative. The Kubernetes API server
silently clamped the negative gracePeriodSeconds to 0, causing pods to
be force-deleted. This violates at-most-one semantics because the old
container may still be running on the instance while a replacement pod
starts elsewhere.

Fix: when the NodeClaim has a TerminationGracePeriod configured, set
the annotation to now()+TGP rather than now(), matching the semantics
of the nodeclaim/lifecycle controller which correctly computes
DeletionTimestamp+TGP. When there is no TGP configured, now() is still
used since there is no grace window to forward-date.

Also add a defensive clamp (>= 0) on gracePeriodSeconds inside
DeleteExpiringPods so that future callers who set a past timestamp
produce a clearly-bounded zero-second grace rather than relying on the
API server's silent clamp.

Signed-off-by: praveen9354 <praveen9354@gmail.com>

Pull Request Pull Request #3038: fix(termination): prevent negative pod grace period when Node Repair triggers

Coverage Stats

18 of 20 new or added lines in 2 files covered. (90.0%)

7 existing lines in 1 file now uncovered.

12523 of 15538 relevant lines covered (80.6%)

118631.75 hits per line