7 Oct
2020
7 Oct
'20
10:57 a.m.
Hello everyone,
I have single Ceph cluster with multiple pools for purposes of
Kubernetes PVs and Virtualization (biggest pool, using snapshots here).
Every client cluster has its own pool with max_bytes and max_objects
quotas in place. When quotas are reached, Ceph halts all write I/O
(which I suppose is OK).
I am trying to setup best possible monitoring to prevent reaching pool
quotas, but I am really struggling to find correct metrics Ceph is
driven by, because almost every command I try gives me different results
(rados df, rbd du, ceph health detail) and I feel like quotas are not
usable for this purpose.
After some fiddling with test pool I have somehow got it into state,
when there are some phantom objects in stats (maybe also written to
OSDs) which cannot be listed, accessed or deleted, but are counted into
pool quotas.
-> ceph version
ceph version 14.2.10 (9f0d3f5a3ce352651da4c2437689144fcbec0131) nautilus
(stable)
I have my test pool here (quotas 4k objects, 15 GiB data):
-> ceph osd pool ls detail | grep mirektest
pool 2 'mirektest' replicated size 3 min_size 2 crush_rule 0 object_hash
rjenkins pg_num 32 pgp_num 32 autoscale_mode warn last_change 21360 lfor
0/10929/10927 flags hashpspool,selfmanaged_snaps max_bytes 16106127360
max_objects 4000 stripe_width 0 target_size_bytes 10737418240
application rbd
Cluster health says pool it getting full:
-> ceph health detail
HEALTH_WARN 1 pools nearfull
POOL_NEAR_FULL 1 pools nearfull
pool 'mirektest' has 3209 objects (max 4000)
pool 'mirektest' has 13 GiB (max 15 GiB)
But pool looks completely empty (no output, so I tried JSON output also):
-> rbd -p mirektest --format json ls
[]
-> rbd -p mirektest --format json du
{"images":[],"total_provisioned_size":0,"total_used_size":0}
I am not using pool snapshots:
-> rados -p mirektest lssnap
0 snaps
But there are some RADOS object in it:
-> rados -p mirektest df
POOL_NAME *USED**OBJECTS*CLONES *COPIES*MISSING_ON_PRIMARY UNFOUND
DEGRADED RD_OPS RD WR_OPS WR USED COMPR UNDER COMPR
mirektest *192 KiB* *3209* 0 *9627* 0
0 0 23998515 192 GiB 1221030 297 GiB 0 B 0 B
Okay, lets try to list them:
-> rados -p mirektest ls --all
rbd_directory
rbd_info
rbd_trash
Thats weird, expected more than 3 objetcs. Let's check it's sizes and
content:
-> rados -p mirektest stat rbd_directory
mirektest/rbd_directory mtime 2020-10-07 11:55:16.000000, size 0
-> rados -p mirektest stat rbd_info
mirektest/rbd_info mtime 2019-12-05 12:11:39.000000, size 19
-> rados -p mirektest get rbd_info -
overwrite validated
-> rados -p mirektest stat rbd_trash
mirektest/rbd_trash mtime 2020-10-07 11:55:17.000000, size 0
Do any of you have some idea what's happening here? I am trying to find
way how to cleanup pool without interrupting existing content in
general. Also I have probably no idea how to replicate this and it's not
the story of every pool in cluster. There is e.g. pool which was used
for Kubernetes but it has no issues:
-> rados df
POOL_NAME USED OBJECTS CLONES COPIES MISSING_ON_PRIMARY
UNFOUND DEGRADED RD_OPS RD WR_OPS WR USED COMPR UNDER
COMPR
pool12 192 KiB 3 0 9
0 0 0 8710688 807 GiB 5490880 576 GiB 0
B 0 B
I don't know how to check other pools, because there is data in them .
How can I rely on quotas when in empty cluster there is already 13GiB
data and 3.2k objects (which btw corresponds because 3200 * 4MiB = 12.8
GiB).
Also is there any way how to calculate same number of objects/bytes Ceph
is using to enforce quotas?
Thanks for you time if you read it until the end, really appreciated
that! :-)
Any idea or hint is welcome.
--
Miroslav Kalina
Systems development specialist
miroslav.kalina(a)livesport.eu
+420 773 071 848
Livesport s.r.o.
Aspira Business Centre
Bucharova 2928/14a, 158 00 Praha 5
www.livesport.eu