Crier failing to report OOMed pods

[BenTheElder]

In prow.k8s.io we're seeing some pods fail to report status beyond "running" that have really been OOMed (if we manually inspect via the build cluster)

See: https://kubernetes.slack.com/archives/C7J9RP96G/p1721926824700569 and other #testing-ops therads

/kind bug

[jihoon-seo]

I guess that you mean crier.. 😊
/retitle Crier failing to report OOMed pods

[dims]

😢

[petr-muller]

I don't think this is crier - that only comes later. I think this is podutils interacting with kubernetes/kubernetes#117793 .

NAME                                   READY   STATUS      RESTARTS   AGE
3b953040-572a-400f-8024-02310b8e54f8   1/2     OOMKilled   0          93m

The above (from the linked Slack thread) indicates one of the two containers in the Pod is still alive - sidecar. That container waits for entrypoint wrapper to write how the actual test process (entrypoints subprocess) ended, but with kubernetes/kubernetes#117793 all processes in the containers get OOMkilled, including entrypoint. In the past only the test process would be killed so entrypoint would be able to collect its exit code and exit normally.

Not sure about what would be the solution yet.

[BenTheElder]

IMHO crier should still eventually report what happened here because it is observing the pod success / fail from the outside, but we see pods being perma reported as running (and if they really are running we should be enforcing a timeout via plank?)

We could either hand off more of this to crier / plank to observe the test container failure, or we could introduce additional heart beating / signaling between the entry point and sidecar?

[petr-muller]

True; you are right that Plank should eventually reap a pod stuck like this - if that does not happen then it's perhaps a separate bug, not sure if plank just uses too soft signals (sidecar iirc ignores some signals in certain state when it thinks it still needs to finish uploading artifacts) or it somehow fails to enforce a timeout entirely.

[BenTheElder]

Left a breadcrumb on kubernetes/kubernetes#117793 because I think it's worth noting when the project breaks itself with breaking changes :+)

This problem is probably going to get worse as cgroup v2 rolls out to the rest of our clusters and more widely in the ecosystem.

[BenTheElder]

https://prow.k8s.io/view/gs/kubernetes-jenkins/logs/ci-kubernetes-e2e-kops-aws-scale-amazonvpc-using-cl2/1818540717593595904

https://gcsweb.k8s.io/gcs/kubernetes-jenkins/logs/ci-kubernetes-e2e-kops-aws-scale-amazonvpc-using-cl2/1818540717593595904/

We actually appear to have uploaded all the artifacts ?? But we're just not recording a finished.json or the test log

[BenTheElder]

I'm seeing a lack of metadata (like the pod json) on pods that received SIGTERM and probably(?) weren't OOMed, at least there's no indication they would be, monitoring suggests they're actually not using much of the memory requested/limited.

https://prow.k8s.io/job-history/gs/kubernetes-jenkins/logs/ci-etcd-robustness-arm64

https://prow.k8s.io/view/gs/kubernetes-jenkins/logs/ci-etcd-robustness-arm64/1829281539263827968

[BenTheElder]

In that case we have some other weirdness: https://kubernetes.slack.com/archives/CCK68P2Q2/p1724976972581789?thread_ts=1724975407.113939&cid=CCK68P2Q2