Hi, this is unexpected, of course, but it can happen if one OSD is full (or also nearfull?). Have you checked 'ceph osd df'? The pg availability has more priority than the placement, so it's possible that during a failure some chunks are recreated on the same OSD or host even if the crush rules shouldn't allow that. Regards, Eugen Zitat von Maks Kowalik <maks_kowalik@poczta.fm>:

Hello,

I have created a small 16pg EC pool with k=4, m=2. Then I applied following crush rule to it:

rule test_ec { id 99 type erasure min_size 5 max_size 6 step set_chooseleaf_tries 5 step set_choose_tries 100 step take default step choose indep 3 type host step chooseleaf indep 2 type osd step emit }

The OSD tree looks as following: -1 43.38448 root default -9 43.38448 region lab1 -7 43.38448 room dc1.lab1 -5 43.38448 rack r1.dc1.lab1 -3 14.44896 host host1.r1.dc1.lab1 6 hdd 3.63689 osd.6 up 1.00000 1.00000 8 hdd 3.63689 osd.8 up 1.00000 1.00000 7 hdd 3.63689 osd.7 up 1.00000 1.00000 11 hdd 3.53830 osd.11 up 1.00000 1.00000 -11 14.44896 host host2.r1.dc1.lab1 4 hdd 3.63689 osd.4 up 1.00000 1.00000 9 hdd 3.63689 osd.9 up 1.00000 1.00000 5 hdd 3.63689 osd.5 up 1.00000 1.00000 10 hdd 3.53830 osd.10 up 1.00000 1.00000 -13 14.48656 host host3.r1.dc1.lab1 0 hdd 3.57590 osd.0 up 1.00000 1.00000 1 hdd 3.63689 osd.1 up 1.00000 1.00000 2 hdd 3.63689 osd.2 up 1.00000 1.00000 3 hdd 3.63689 osd.3 up 1.00000 1.00000

My expectation was that each host will contain 2 shards of any PG of the pool.

When I dumped PGs, it was true, but one group is placed on OSDs 0,2,3 which will cause downtime in case of host3 failure. root@host1:~/mkw # ceph pg dump|grep "^66\."|awk '{print $17}' dumped all [4,5,7,6,1,2]

[8,11,9,3,0,2] <<< - this one is problematic

[6,7,10,9,2,0] [2,3,7,6,5,9] [7,8,10,5,3,1] [4,5,8,6,0,2] [7,11,9,4,1,2] [5,9,0,2,7,11] [9,5,3,1,7,8] [8,11,2,0,5,9] [2,0,8,6,10,9] [3,2,5,9,7,11] [6,7,9,5,1,2] [10,5,1,3,11,8] [4,5,7,8,2,0] [7,8,3,2,9,10]

Is there a way to ensure that host failure is not disruptive to the cluster?

During the experiment I used info from this thread: http://lists.ceph.com/pipermail/ceph-users-ceph.com/2018-October/030227.html

Kind regards,

Maks Kowalik _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io