[LU-17531] 'lfs mirror extend' should succeed with stale mirror on open-for-write files

• Type: Improvement
• Resolution: Unresolved
• Priority: Minor
• Affects Version/s: Lustre 2.14.0, Lustre 2.16.0
• Labels: FLR flr-improvement hotpools medium usability
• Severity: 3
• Epic Link: Hot Pools backlog
• Assignee: Zhenyu Xu
• Reporter: Andreas Dilger

Description

Rather than completely failing the "lfs mirror extend" command when the MDS finds the mirror version to be different, it would be better for the MDS to allow the layout add to complete, but mark the new mirror LCME_FL_STALE immediately.

This would allow adding redundancy to files that are continually being written. With "--block" we shouldn't need this, but it isn't always nice to the application to block IO for 100s while a large file is being completely mirrored.

This also potentially allows future optimizing the resync of the stale mirror instead of having to create it again. For example, if the mirror is extended incrementally by component and the stale flag is only added to mirror components that were actually modified during this operation then it would mean less work to resync later.

This should add an option to "lfs mirror extend" to enable/disable this behavior, maybe "--allow-stale" and "--deny-stale"?

I think the default should be to enable this new behavior, but it should be possible to disable it on the MDS if it is causing problems. I think the affected logic will be in the LOD layer, so a parameter like "lod.*.enable_mirror_extend_stale" or similar should be added.

Issue Links

• is related to: LU-12649 Tracker for ongoing FLR improvements
• is related to: LU-18332 lfs migrate is not setting the mirror count
• is related to: LU-14554 Improve usability of "lfs mirror" commands
• is related to: LU-13668 open-for-read should not conflict with mirror creation
• is related to: LU-18746 'lfs mirror extend' should allow creating a new STALE mirror