Only rank=0 metrics endpoint should be fetched in autoscaler if podGroupSize is configured

[Jeffwan]

🐛 Describe the bug

if we use podGroupSize, the httpserver will be only launched in the rank=0 server, rest pods are pure GPU workers. In that case, we should not fetch the endpoint. this is kind of missing in autoscaler now.

Steps to Reproduce

PodGroupSize != 1

Expected behavior

autoscaler should only fetch the pods with rank=0.

Environment

nightly

[scarlet25151]

Hi @Jeffwan, just want to double confirm that if all server would launched on pod with pod index 0 in the podset? Is there any other pod selector?

[Jeffwan]

@scarlet25151

aibrix/pkg/controller/constants/stormservice.go

Line 29 in 766375d

PodGroupIndexLabelKey = "stormservice.orchestration.aibrix.ai/pod-group-index"