On Wed, 24 Jan 2018, Piotr Dałek wrote: > On 18-01-24 10:54 AM, Vincent Godin wrote: > > I have a cluster in recovery state and i notice that few pgs are only > > hold bye one OSD and a lot hold by two OSDs (pool configured with size > > =3 & min=1). After few days, i'm surprised to see that the few pgs > > hold only by one OSD are still there. It seems there is no high > > priority for pg in critical state (1 on 3) compare to those in warn > > state (2 on 3). Am i wrong or could it be a near feature ? > > My Cluster is in Jewel 10.2.6 > > http://docs.ceph.com/docs/master/release-notes/#v10-2-7-jewel > https://github.com/ceph/ceph/pull/13232 might be closest to what you need. > You'd still need to bump up min_size to 2, though. Note that there is also an issue in luminous where a lower-priority recovery may already be in progress when a higher-priority PG becomes ready to recovery, the low-priority PG is not preempted. This is fixed in master for mimic but the backport is nontrivial. sage
