search

gardener / autoscaler

Customised fork of cluster-autoscaler to support machine-controller-manager
Apache License 2.0
16 stars 25 forks source link

Scale from Zero doesn't work in case of pods with PVs #245

Closed himanshu-kun closed 11 months ago

himanshu-kun commented 1 year ago

What happened:

Customer had a pod with PV and a node group which scales from zero. When the customer hibernates the cluster (means nodegrp scaled down to zero) and then wake-it up again (means pod gets deployed , and old PVC, PV already present) then the scale-up from zero doesn't happen.

This behaviour is not observed when pod uses a new PVC (which doesn't have an associated PV)

What you expected to happen: Scale from zero to work everytime

How to reproduce it (as minimally and precisely as possible): explained above

Anything else we need to know: This is happening because there are topology.ebs.csi.. labels which are placed by KCM on the node and the affinity is placed on the PV by the csi-driver for zone pinning. But the autoscaler is not aware of these labels, and the labels vary from provider to provider. This is why scale from 1 scenario passes while scale from zero fails

The issue has been fixed for AWS , Azure and GCP , but is left for other providers listed below

Environment:

CA 1.27.0 and less

himanshu-kun commented 1 year ago

/priority 2 /assign @himanshu-kun

important to be implemented before https://github.com/gardener/gardener/issues/7857

briantopping commented 1 year ago

Hi I think we have answered relevant questions for vSphere. The documentation misrepresents that the vSphere backend should be a specific version. As maintainer of that plugin, this doesn't need to be a concern. 🤓

ashwani2k commented 1 year ago

Can we update the issue with the corresponding provider specific PRs to make it easy to track them later when required.

elankath commented 1 year ago

Updating issue with relevant PR's

himanshu-kun commented 11 months ago

/close as done