Skip to content

Commit c09cda5

Browse files
committed
fix(e2e): give the kafka teardown delete its own timeout
The verify-pvc-reclaim-on-delete step deletes the Kafka CR and waits on its finalizer, which unwinds a serial stack: the apps controller removes the backing HelmRelease, Flux uninstalls the chart, and the Strimzi operator terminates the broker and ZooKeeper StatefulSets and reclaims their PVCs (deleteClaim). That is bounded, converging work, not a hang. On the reported run the CR and its HelmRelease were both gone ~5m13s after the delete, landing right on the suite-default 5m delete budget, so the wait flakes on the boundary. That default was sized for finalizer-draining CRs elsewhere (VMDisk, FoundationDB, Tenant), so scope the fix to this step rather than widen the global for everyone. 8m sits above the observed cost with headroom and stays finite, so a genuine stuck uninstall still fails loudly. Fixes #4260 Signed-off-by: Alexey Artamonov <[email protected]>
1 parent 070dcb1 commit c09cda5

1 file changed

Lines changed: 14 additions & 0 deletions

File tree

‎hack/e2e-chainsaw/kafka/chainsaw-test.yaml‎

Lines changed: 14 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -94,7 +94,21 @@ spec:
9494

9595
- name: verify-pvc-reclaim-on-delete
9696
try:
97+
# Kafka teardown unwinds a whole stack behind the CR finalizer that this
98+
# delete waits on: the apps controller removes the backing HelmRelease, Flux
99+
# uninstalls the chart, and the Strimzi operator then terminates the broker
100+
# and ZooKeeper StatefulSets and reclaims their PVCs (deleteClaim: true).
101+
# That is serial, bounded work, not a hang — the create side already budgets
102+
# 6m for the HelmRelease and 5m for Kafka readiness, and a coming-down stack
103+
# costs about the same. On the observed run the CR and its HelmRelease were
104+
# both gone ~5m13s after the delete, right on the suite-default 5m budget, so
105+
# the wait flakes on the boundary. That default was sized for
106+
# finalizer-draining CRs elsewhere (VMDisk, FoundationDB, Tenant), so give
107+
# this step its own ceiling with headroom above the observed cost rather than
108+
# widening the global for everyone. 8m stays finite, so a genuine teardown
109+
# regression (a stuck uninstall that never converges) still fails loudly.
97110
- delete:
111+
timeout: 8m
98112
ref:
99113
apiVersion: apps.cozystack.io/v1alpha1
100114
kind: Kafka

0 commit comments

Comments
 (0)