On Wed, Dec 9, 2015 at 1:25 PM, Mykola Dvornik <[email protected]> wrote: > Hi Jan, > > Thanks for the reply. I see your point about replicas. However my motivation > was a bit different. > > Consider some given amount of objects that are stored in the metadata pool. > If I understood correctly ceph data placement approach, the number of > objects per PG should decrease with the amount of PGs per pool. > > So my concern is that in catastrophic event of some PG(s) being lost I will > loose more objects if the amount of PGs per pool is small. At the same time > I don't want to have too few objects per PG to keep things disk IO, but not > CPU bounded.
If you are especially concerned about triple-failures (i.e. permanent PG loss), I would suggest you look at doing things like a size=4 pool for your metadata (maybe on SSDs). You could also look at simply segregating your size=3 metadata on to separate spinning drives, so that these comparatively less loaded OSDs will be able to undergo recovery faster in the event of a failure than an ordinary data drive that's full of terabytes of data, and have a lower probability of a triple failure. John _______________________________________________ ceph-users mailing list [email protected] http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
