vllm-project / vllm-project/aibrix
StormService not Rollout unscheduled pods firstly
- Dominant language
- Go
- Stars
- 5.1k
- Forks
- 697
- Avg merge
- 1d 19h
- Merged PRs (30d)
- 104
Description
Recently I use StormService in our production K8S, I found the Rollout pods rollback does not prioritize rolling back unscheduled pods, but rolls
back pods in sequence.
For example, I apply 6 replicas with nodeSelector app=aibrix, pod-0,3,4,5 is scheduled, 1 and 2 is not scheduled, because of only 4 GPU nodes with label app=aibrix are available.
Then I unset label of 2 nodes for other usage temporarily. When I rolleout 6 replicas to 4 replicas, StormService Controller will roll back pod-5 and pod-4, the unscheduled pod-1 and pod-2 is not deleted.
I expect the unscheduled pods can be delete firstly. I want to ask, delete the pods in order is how is the design, or delete the unscheduled pods firstly are taken into account in the future?
Contributor guide
Assessment
This issue has not been assessed yet.