Are you referring to manual roll out of ceph osd crush tunables optimal ( http://lists.ceph.com/pipermail/ceph-users-ceph.com/2014-July/041499.html) ?
Can you include a copy of ceph osd tree and your crushmap? Were the OSDs flapping when you did those queries? pg 6.19 says [0,1] is acting, but only [0] is up. pg 6.0 says [1] is acting and [1] is up. It seems strange that different PGs have a different idea of which OSDs are up. The web site hasn't been updated to state what "undersized" means, but the code says that means there are fewer replicas than requested. That's different from degraded, which is when not all of the replicas are available. I'm not really sure what that means in practice though. If I had to guess, I say the crushmap is bad maybe? On Wed, Nov 5, 2014 at 6:35 AM, Andrey Korolyov <[email protected]> wrote: > Hello, > > after doing a single-step transition, the test cluster is hanging in > unclean state, both before and after crush tunables adjustment: > > status: > http://xdel.ru/downloads/transition-stuck/cephstatus.txt > > osd dump: > http://xdel.ru/downloads/transition-stuck/cephosd.txt > > query for a single pg in active+remapped state: > http://xdel.ru/downloads/transition-stuck/remappedpg.txt > > query for a single pg in active+undersized+degraded state: > http://xdel.ru/downloads/transition-stuck/degradedpg.txt > > As one can see there is empty value for backfill_targets for both sets > of pg which is clearly indicates some problem with placement > calculation (two-node two-osd cluster should have enough targets for > backfilling in this case). > _______________________________________________ > ceph-users mailing list > [email protected] > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >
_______________________________________________ ceph-users mailing list [email protected] http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
