[Bug] Kuberay autoscaler should use numOfHosts to calculate max workers

[richardsliu]

Search before asking

• I searched the issues and found no similar issues.

KubeRay Component

ray-operator

What happened + What you expected to happen

Kuberay autoscaler currently uses the maxReplicas field in the worker group spec to determine maxWorkers: https://github.com/ray-project/ray/blob/master/python/ray/autoscaler/_private/kuberay/autoscaling_config.py#L197

This is used by the Ray autoscaler to determine if workers need to be terminated: https://github.com/ray-project/ray/blob/master/python/ray/autoscaler/_private/autoscaler.py#L1004

The maximum number of pods should be maxReplicas * numOfHosts, otherwise when numOfHosts > 1 this could result in workers getting terminated unintentionally.

Reproduction script

n/a

Anything else

No response

Are you willing to submit a PR?

• Yes I am willing to submit a PR!

[ryanaoleary]

Related PR that would close this issue: ray-project/ray#48212