pg inconsistent state is lost after the primary osd restart
Steps to reproduce:
- Create a pool (either replicated or erasure)
- Introduce an inconsistency (e.g. put an object and then remove one replica or shard)
- Run pg deep-scrub to detect the inconsistency and the pg enter 'active+clean+inconsistent' state
- Restart the primary osd.
- After the osd is up again and the pg is peered it is in 'active+clean' state.
The expected result: the pg 'inconsistent' state is not lost after the primary osd restart.
#4 Updated by Mykola Golub over 2 years ago
yite gu wrote:
What is the way you remove replica?
In my case it was filestore so I just remove the file on the fs. With bluestore I would use ceph-objectstore-tool for this. But I don't think the nature of inconsistency is important here. The first time I noticed the issue was when I was investigating an inconsistency introduced by invalid hinfo, and then later when checking it was easily reproducible I was introducing an inconsistency by just removing a replica.