vllm-project / vllm-project/aibrix
rayfleet scale down support ways to determine the scaling-down priority.
Open
area/orchestration
kind/bug
- Dominant language
- Go
- Stars
- 5.1k
- Forks
- 694
- PR merge metrics
- PR metrics pending
Description
### 🚀 Feature Description and Motivation
Currently, when scaling down a RayClusterReplicaSet, it directly removes the first N rayclusters in the list. It doesn't support scaling down specific replicas, which might result in deleting ready replicas while keeping unready ones, making the whole service unavailable. We need to support ways to determine the scaling-down priority of rayclusters, like sorting by deletion cost.
### Use Case
When reducing the number of replicas in rayfleet, the cluster list should be sorted in reverse order according to the deletion cost.
### Proposed Solution
_No response_
Contributor guide
Assessment
This issue has not been assessed yet.