Re: jewel ceph has PG mapped always to the same OSD's

[Konstantin Danilov <kdanilov@xxxxxxxxxxxx>]

David,

> What happens when you deep-scrub this PG?
we haven't try to deep-scrub it, will try.

> What do the OSD logs show for any lines involving the problem PGs? 
Nothing special were logged about this particular osd, except that it's degraded.
Yet osd consume quite a lot portion of its CPU time in snappy/leveldb/jemalloc libs.
In logs there a lot of messages from leveldb about moving data between levels.
Needles to mention that this PG is from RGW index bucket, so it's metadata only

and get a relatively hight load. Yet not we have 3 PG with the same
behavior from rgw data pool ()cluster have almost all data in RGW

> Was anything happening on your cluster just before this started happening at first?
Cluster gets many updates in a week before issue, but nothing particularly noticeable.

SSD OSD get's split in two, about 10% of OSD were removed. Some networking issues
appears.

Thanks

On Fri, Apr 6, 2018 at 10:07 PM, David Turner <drakonstein@xxxxxxxxx> wrote:

What happens when you deep-scrub this PG?  What do the OSD logs show for any lines involving the problem PGs?  Was anything happening on your cluster just before this started happening at first?

On Fri, Apr 6, 2018 at 2:29 PM Konstantin Danilov <kdanilov@xxxxxxxxxxxx> wrote:

Hi all, we have a strange issue on one cluster.

One PG is mapped to the particular set of OSD, say X,Y and Z doesn't
matter what how
we change crush map.
The whole picture is next:

* This is 10.2.7 ceph version, all monitors and osd's have the same version
* One  PG eventually get into 'active+degraded+incomplete' state. It
was active+clean for a long time
and already has some data. We can't detect the event, which leads it
to this state. Probably it's
happened after some osd was removed from the cluster
* This PG has all 3 required OSD up and running, and all of them
online (pool_sz=3, min_pool_sz=2)
* All requests to pg stack forever, historic_ops shows that it waiting
on "waiting_for_degraded_pg"
* ceph pg query hangs forever
* We can't copy data from another pool as well - copying process hangs
and that fails with
(34) Numerical result out of range
 * We was trying to restart osd's, nodes, mon's with no effects
* Eventually we found that shutting down osd Z(not primary) does solve
the issue, but
only before ceph set this osd out. If we trying to change the weight
of this osd or remove it from cluster problem appears again. Cluster
is working only while osd Z is down and not out and has the default
weight
* Then we have found that doesn't matter what we are doing with crushmap -
osdmaptool --test-map-pgs-dump always put this PG to the same set of
osd - [X, Y] (in this osdmap Z is already down). We updating crush map
to remove nodes with OSD X,Y and Z completely out of it, compile it,
import it back to osdmap and run osdmaptool and always get the same
results
* After several nodes restart and setting osd Z down, but no out we
are now have 3 more PG with the same behaviour, but 'pined' to another
osd's
* We have run osdmaptool from luminous ceph to check if upmap
extension is somehow getting into this osd map - it is not.

So this is where we are now. Have anyone seen something like this? Any
ideas are welcome. Thanks


--
Kostiantyn Danilov

_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com

--

Kostiantyn Danilov aka koder.ua
Principal software engineer, Mirantis

skype:koder.ua
http://koder-ua.blogspot.com/
http://mirantis.com

_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com