Evict slow store scheduler might not be recoverd forever
Open
component/mcs
nextgen
type/bug
- Dominant language
- Go
- Stars
- 1.2k
- Forks
- 783
- Avg merge
- 5d 21h
- Merged PRs (30d)
- 36
Description
## Bug Report
In the microservice environment, we use a memory storage in the scheduling service. But the evict slow store scheduler need to persist the state of evicted store. After the PD leader changes, the persisted state will lose. The region leader can be transferred again.
Another issue is if we changes the config such as the batch/recovery time when some store is in evicted state, the scheduling service will update that config with evicted stores information. Under the circumstances, the slow store scheduler cloud skip cleaning the evicted store state forever even if it exceeds the recovery time.
Contributor guide
Assessment
This issue has not been assessed yet.