Join GitHub today
GitHub is home to over 28 million developers working together to host and review code, manage projects, and build software together.Sign up
osd: process deletes during recovery instead of peering #15952
With a large number of deletes in a client workload, performing them during peering can easily saturate a disk and cause very high latency, since it does not go through the op queue or do any batching.
To fix this, make recovery handle these deletes instead of performing them during peering.
This looks good! I think we need the min_epoch pieces or else we may get odd ordering issues between messages (everything else sets min_epoch and if these don't they may get blocked when others don't). (Also it's faster.)