Fixing a kubelet memory leak in Kubernetes 1.36

https://heyoncall.com/blog/fixing-kubernetes-kubelet-memory-leak

by compumike • 22 days ago

79 14 22 days ago

compumike5• 22 days ago

Author here! If you're running a Kubernetes cluster, I recommend you check `kubectl version` and see if you're running "Server Version: v1.36.[0,1,2]". If so, you may want to use the one-liner at the end of the article to check your "process_resident_memory_bytes" on each node, and consider restarting kubelet as a temporary workaround to tame the memory leak until v1.36.3 is released.

rirze• 21 days ago

Very cool. It's often daunting to contribute to such a well-established and recognizable project, but this is exactly how it should work.

CamouflagedKiwi• 21 days ago

Nice find.

Can't help but feel this is one of the subtle traps hidden beneath the advice that contexts aren't supposed to be stored. I know it's not always that easy, of course.

compumike5• 21 days ago

Thanks. I know there's a `go vet` tool that's run as part of Kubernetes CI, and one of its checks is:

  lostcancel: check cancel func returned by context.WithCancel is called

I'm not 100% sure why `go vet` didn't catch this issue, but storing the cancelFn in the struct is probably part of the reason. Any Go experts know if that's the case?

cyberax2• 21 days ago

The cancel function escapes the function body, so static analysis can't detect it. There's another lint for that (containedctx), but I think it's off in K8s.

This is a serious tripping point with Go. There's no way to express: "this is a root context that I _want_ to store and only use to create derived contexts". Goroutines are also a source of problems, you can't easily say "I'm passing the ownership of this context to a goroutine".

fsuts• 21 days ago

Not all heroes wear capes! Well done

__turbobrew__• 21 days ago

A good reason to health check the kubelet process and restart it when the checks fail.

compumike5• 21 days ago

What kind of health checks? In my case, the kubelet process was staying alive and responsive to queries, I believe due to:

  # cat /proc/$(pgrep kubelet)/oom_score_adj
  -999
  
  (from OOMScoreAdjust=-999 in /etc/systemd/system/kubelet.service)

With this score, the Linux OOM killer wouldn't touch it, but any of my Pods were fair game.

nijave• 21 days ago

At the metrics level, you can compare old vs new release. Have been bitten before by resource requirements dramatically change (regardless of whether it's a bug or functionality change)

jeffbee• 21 days ago

Given that this happens on the main loop, one wonders how this escaped release quality checks. Is there another contributing cause that narrows the impact?

compumike5• 21 days ago

I don’t see any other impact-narrowing criteria.

The default kubelet `syncFrequency` is 1 minute (per Pod). (There can be additional event-driven ones, but this is a floor.) Back of the envelope: 30 Pods * 1440 * 7 = 300k calls to startPodSync in a week.

I’d guess a lot of production clusters don’t run the latest release of Kubernetes, and 1.36 was only released in late April, less than two months before I reported the issue.

Given that the issue shows up as linear memory growth on the order of maybe 1 GiB/month per node, there really is a time-based discovery process here. (Could be artificially accelerated for a CI test, I imagine!) And any restarted nodes or clusters that upgraded point releases would restart that clock.

The only unusual contributing cause that made the issue more visible to me was trying to run everything on a tiny 2 GiB RAM node!

lanycrost• 21 days ago

Nice finding, golang and kubernetes toolset is always surprisingly enough and friendly for doing such debuggings. Currently me to running this version so kudos to you!

Fixing a kubelet memory leak in Kubernetes 1.36

Related Stories