Actions
Feature #4485
openImprove "needsrecover" handling
Status:
New
Priority:
Normal
Assignee:
-
Category:
Performance/Resource Usage
Target version:
-
% Done:
0%
Source:
Development
Tags:
Backport:
Reviewed:
Affected Versions:
Component(FS):
Common/Protocol, MDS
Labels (FS):
Pull request ID:
Description
Jim Schutt reported issues on the mailing list1 with slow stats that turned out to be due to inodes with the "needsrecover" flag being recovered inline with the stats occurring. This despite the fact that the inodes had been so marked for hours or days prior, following a flapping MDS and client session reconnects. Examine the current recovery framework and identify ways to optimize it.
Possibilities:
1) do we not have an async recovery? I thought we did; if we don't we should.
2) Perhaps it was due to lock states. Clean it up so the MDS more aggressively moves needsrecover files into states where it can do the recovery.
[1]: Starting approximately here: http://www.spinics.net/lists/ceph-devel/msg13217.html
Actions