Hi Michel,
Are you sure there isn't a hardware problem with the disk? E.g.
maybe you have SCSI timeouts in dmesg or high ioutil with iostat?
Anyway I don't think there's a big risk related to draining and
stopping the osd. Just consider this a disk failure, which can
happen at any time anyway.
Start by marking it out. If there are still too many slow requests
or laggy PGs, try setting primary affinity to zero.
And if that still doesn't work, I wouldn't hesitate to stop that
sick osd so objects backfill from the replicas
(We had a somewhat similar issue today, btw .. some brand of SSDs
occasionally hangs IO across a whole SCSI bus when failing. Stopping
the osd revives the rest of the disks on the box).
Cheers, Dan
On Sun, Oct 16, 2022, 22:08 Michel Jouvin <jouvin@xxxxxxxxxxxx> wrote:
Hi,
We have a production cluster made of 12 OSD servers with 16 OSD each
(all the same HW) which has been running fine for 5 years (initially
installed with Luminous) and which has been running Octopus
(15.2.16)
for 1 year and was recently upgraded to 15.2.17 (1 week before the
problem started but doesn't seem to be linked with this upgrade).
Since
beginning of October, we started to see PGs in state
"active+laggy" and
slow requests always related to the same OSD and looking at its
log, we
saw "log_latency_fn slow" messages. There was no disk error
logged in
any system log file. Restarting the OSD didn't really help but no
functionnal problems were seen.
Looking again at the problem in the last days, we saw that the
cluster
was in HEALTH_WARN state because several PGs were not
deep-scrubbed in
time. In the logs we saw also (but may be we just missed them
initially)
"heartbeat_map is_healthy 'OSD::osd_op_tp thread...' had timed out
after
15" messages. This number increased days after days and is now
almost 3
times the number of PGs hosted by the laggy OSD (despite hundreds of
deep scrubs running successfully, the cluster has 4297 PGs). It
seems
that in the list we find all PGs that have a replica (all the
pools are
with 3 replica, no EC) on the laggy OSD. We confirmed that there
is no
detected disk error in the system.
Today we restarted the server hosted this OSD, without much hope. It
didn't help and the same OSD (and only this one) continues to have
the
same problem. In addition to the messages mentioned, the admin
socket
for this OSD became unresponsive: despite command being executed
(see
below), it was not returning in a decent amount of times (several
minutes).
As the OSD RocksDB have probably never been compacted, we decided to
compact the laggy OSD. Despite the "ceph tell osd.10 compact" never
returned (it was killed after a few hours as the OSD has been marked
down during a few seconds), the compaction started and lasted ~5
hours... but completed successfully. But the only improvement that
was
seen after the compaction was that the admin socket is now
responsive
(despite a bit slow). The messages about log_latency_fn and
heartbeat_map are still present (and frequent) and the deep scrubs
are
still blocked.
We are looking for advices on what to do to fix this issue. We'd
in mind
to stop this OSD, zap it and resintall it but we are worrying it
may be
risky to do this with an OSD that has not been deep scrubbed for a
long
time. And we are sure there is a better solution! Understanding the
cause would be a much better approach!
Thanks in advance for any help. Best regards,
Michel
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
