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
os/filestore/HashIndex: be loud about splits #12421
This ML thread "[ceph-users] filestore_split_multiple hardcoded maximum" has coincided with ~1 week of slow requests on one of our clusters. After much debugging we finally tracked it down to filestore splitting.
It would have been much easier to understand the root cause if filestore splitting was verbose. Splitting is a rare but rather important event -- the OSD should be super verbose when it happens. (Otherwise, it simply looks like a filestore write is taking many 10s of seconds, which seems absurd).
So I'm sending this patch for comments. I'm not sure if the dout is needed in the _create function, and I didn't make the merges verbose yet (which I guess is also needed?)
If this makes it to master, I'd appreciate a backport to jewel. (filestore will lose relevance anyway in L and beyond).
@cernceph Could you open a tracker issue for this? You can flag the issue for jewel backport while doing that, just fill in Backport: jewel.
If you open the tracker, please also add a line
changed the title from
RFC: os/filestore/HashIndex: be loud about splits
os/filestore/HashIndex: be loud about splits
Feb 28, 2017
same job passed on manual rerun !
and confirmed pass
TestLibRBD.FlattenNoEmptyObjects failures unrelated, also on master per @jasondillaman