On Wed, Nov 16, 2016 at 1:16 AM, Patrick Donnelly <pdonnell@xxxxxxxxxx> wrote: > On Tue, Nov 15, 2016 at 8:40 AM, Hauke Homburg <hhomburg@xxxxxxxxxxxxxx> wrote: >> In the last weeks we enabled for testing the dir fragmentation. The Resultat >> is that we have sometimes error messages with rsync with unlink and no-space >> left on device. > > Enabling directory fragmentation would not cause the unlink and ENOSPC > errors. Failure to unlink is caused by the stray directories on the > MDS growing too large. The only current solution is to wait for the > MDS to eventually purge the stray directory entries. Retry the unlink > as necessary. [The other workaround is to increase > mds_bal_fragment_size_max [1] which is not recommended.] > > Directory fragmentation is not yet considered stable so beware > potential issues including data loss. However, fragmentation will > allow your directories to grow to unbounded size. This includes the > stray directories which would permit unlink to avoid this issue. The last part isn't quite right, we currently don't fragment strays[1]. Unfortunately anyone who uses directory fragmentation to create a super-big directory could still have issues when unlinking it. However, there are 10x stray directories and removed items are spread between them, so you should be able to handle deleting a directory 10x the limit on the size of a stray dir. As Patrick says, if you do run into this case then you can always get around it by just repeating your deletes until they go through. John 1. Rather than lifting this restriction in Luminous, we're going to be changing the way deletion works to avoid so many things going into strays to begin with. >> Does anyone have a timeline for the testing dir frag mds? > > Directory fragmentation is on track to be stable for the Luminous release. > > [1] https://github.com/ceph/ceph/pull/9789 > > -- > Patrick Donnelly > _______________________________________________ > ceph-users mailing list > ceph-users@xxxxxxxxxxxxxx > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |