Questions for configurations using Helm Chart

How severe does this issue affect your experience of using Ray?

  • Low: It annoys or frustrates me for a moment.

I noticed the limits CPU and requests CPU can be set on the values.yaml (kuberay/values.yaml at master · ray-project/kuberay · GitHub). I know the legacy K8s operator can only set up one value(both for limits and requests) and it only can be an integer.

For the KubeRay:

  1. For CPU, can it be set as non-integer, such as 1500m?
  2. For CPU, is it best practice if I set two different values for limits and requests?
  3. For Memory, is it best practice if I set two different values for limits and requests?
  4. For Memory, is it ok if I don’t set up the limits and requests? This means the pod can take all memory on the node if the Ray program attempt to.

Related topic: KubeRay cluster config questions

From some discussion on Slack, I think the answers for 2 and 3 would be No.

I think the way that guarantees the most reliability with Ray is actually to set limits=requests for both memory and CPU.

Ref: Slack

I recommend setting requests and limits equal for both memory and cpu. I wouldn’t omit requests or limits.

I wouldn’t recommend non-integer values for CPU.

https://docs.ray.io/en/latest/cluster/kubernetes/user-guides/config.html#resources