2

The metrics-server-v0.3.6 deployment that is installed into GKE clusters by default is using an outdated version of addon-resizer (addon-resizer:1.8.11-gke.0) that causes CPU utilization and throttling issues. See here

The cluster running is on the latest version of the regular release branch (1.20.8-gke.900), when editing the metric-server workload deployment and changing the version of addon-resizer to addon-resizer-1.8.14 the deployment reverts back to the old version that is managed by google.

Is there a way to change the image that the metrics-server-nanny container deployment is using permanently or is there a way to update the entire metrics-server deployment so that the containers that are running are on an update version?

Luke Rixson
  • 607
  • 5
  • 20
  • Not totally sure how much control (if any) we have in any kube-system ns components. There must be an admission controller which rolls back to the default image as you mentioned. It looks like the addon-resizer got updated to 1.8.14 in the latest Kubernetes release, so this might be a matter of checking with Google Support how soon this will be available in my opinion. – CaioT Aug 06 '21 at 13:13

1 Answers1

5

You cannot update those deployments. Any updates you make will be reverted to the original deployment spec.

Looks like the "fix" will be released with 1.22 and initially backported to 1.21 as well. It will likely make it back to 1.19+ eventually.

Gari Singh
  • 11,418
  • 2
  • 18
  • 41
  • If you don't have a ticket open with GCP support, you should do so as then you'll be able to track this. – Gari Singh Aug 06 '21 at 13:23
  • Thanks very much, I had a hunch this was the case but wasn't sure if there was some workaround. It is not worth updating the whole cluster for this issue and isn't causing any real issues so I will wait for the backport. Thanks again! – Luke Rixson Aug 06 '21 at 13:35