statefulset: make pod restart detection idempotent #1133
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Pods can get stuck with ClusterUpdate condition set but never restarted when the status.restarting flag is not persisted due to controller restarts, stale status reads, or status update conflicts with concurrent reportStatus() calls. The controller relied solely on status.restarting to decide if a rolling update should continue, but this breaks idempotency: a fresh reconcile with stale cached status would skip the restart even though pods still had the ClusterUpdate condition set. Fix by checking actual pod state (ClusterUpdate condition) as the authoritative source of truth in shouldUpdate(). If any pods need restart but status flag is false, we now detect this and continue the rolling update, re-establishing the status flag and completing the restart. This makes the controller self-healing: pod conditions are the real state, status flags are just derived state that gets recomputed if lost.
birdayz
requested review from
RafalKorepta,
andrewstucki,
chrisseto and
gene-redpanda
as code owners
RafalKorepta
approved these changes
Oct 9, 2025
Contributor
RafalKorepta
left a comment
RafalKorepta
left a comment
There was a problem hiding this comment.
LGTM, I would highly recommend to create unit test.
Contributor
Author
|
Will add test and fix compile errors |
This test verifies that the controller correctly detects pods with ClusterUpdate condition even when the status cache is stale. It covers the idempotency fix where shouldUpdate() now checks actual pod conditions as the source of truth, not just cached status flags. The test simulates the ~1% case where a fresh reconcile sees stale status: - Pods have ClusterUpdate condition set (authoritative state) - Status flags say restarting=false (stale cached state) - shouldUpdate() must still detect this and return nodePoolRestarting=true Also fixes compilation errors in existing tests by adding context parameter and fake Kubernetes client where needed.
Contributor
|
@RafalKorepta didn't you recently fix this or was it a similar but distinct issue? @birdayz does this need to be backported to v25.1 or is cloud officially on v25.2 beta builds? |
Contributor
Cloud will be on v25.2 just after https://github.com/redpanda-data/cloudv2/pull/23422 which is in the merge queue as of writing.
#1075 is different problem, but touches rolling upgrade process too. |
Contributor
Author
|
i think we need a backport, but rafal knows better |
Contributor
|
I suggest to open new PR in cloud repo that includes this fix. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Problem
Pods get stuck with ClusterUpdate condition set but never restarted. The
controller relied on status.restarting flag which can be lost due to
controller restarts or stale cache reads (~1% of reconciles in distributed
systems). When a fresh reconcile sees restarting=false but pods still have
the condition, it exits early and never completes the restart.
Solution
Check actual pod state (ClusterUpdate condition) as source of truth. If pods
need restart but status flag says otherwise, trust the pods and continue the
rolling update. The status flag gets recomputed from reality.
Pod conditions are the real state. Status flags are just cached hints that
get thrown away and rebuilt if they go stale. This is how distributed systems
should work.