Hi guys
Thanks for the suggestions, we'll do the offline compaction and see how
big an impact it will have.
Even if compact-on-iteration should take care of it, doing it offline
should avoid I/O problems during the compaction, correct? That's why
offline is preferred to online?
Mvh.
Torkil
On 22-03-2024 16:47, Joshua Baergen wrote:
Personally, I don't think the compaction is actually required. Reef
has compact-on-iteration enabled, which should take care of this
automatically. We see this sort of delay pretty often during PG
cleaning, at the end of a PG being cleaned, when the PG has a high
count of objects, whether or not OSD compaction has been keeping up
with tombstones. It's unfortunately just something to ride through
these days until backfill completes.
https://github.com/ceph/ceph/pull/49438 is a recent attempt to improve
things in this area, but I'm not sure whether it would eliminate this
issue. We've considered going to higher PG counts (and thus fewer
objects per PG) as a possible mitigation as well.
Josh
On Fri, Mar 22, 2024 at 2:59 AM Alexander E. Patrakov
<patrakov@xxxxxxxxx> wrote:
Hello Torkil,
The easiest way (in my opinion) to perform offline compaction is a bit
different than what Igor suggested. We had a prior off-list
conversation indicating that the results would be equivalent.
1. ceph config set osd osd_compact_on_start true
2. Restart the OSD that you want to compact (or the whole host at
once, if you want to compact the whole host and your failure domain
allows for that)
3. ceph config set osd osd_compact_on_start false
The OSD will restart, but will not show as "up" until the compaction
process completes. In your case, I would expect it to take up to 40
minutes.
On Fri, Mar 22, 2024 at 3:46 PM Torkil Svensgaard <torkil@xxxxxxxx> wrote:
On 22-03-2024 08:38, Igor Fedotov wrote:
Hi Torkil,
Hi Igor
highly likely you're facing a well known issue with RocksDB performance
drop after bulk data removal. The latter might occur at source OSDs
after PG migration completion.
Aha, thanks.
You might want to use DB compaction (preferably offline one using ceph-
kvstore-tool) to get OSD out of this "degraded" state or as a preventive
measure. I'd recommend to do that for all the OSDs right now. And once
again after rebalancing is completed. This should improve things but
unfortunately no 100% guarantee.
Why is offline preferred? With offline the easiest way would be
something like stop all OSDs one host at a time and run a loop over
/var/lib/ceph/$id/osd.*?
Also curious if you have DB/WAL on fast (SSD or NVMe) drives? This might
be crucial..
We do, 22 HDDs and 2 DB/WAL NVMes pr host.
Thanks.
Mvh.
Torkil
Thanks,
Igor
On 3/22/2024 9:59 AM, Torkil Svensgaard wrote:
Good morning,
Cephadm Reef 18.2.1. We recently added 4 hosts and changed a failure
domain from host to datacenter which is the reason for the large
misplaced percentage.
We were seeing some pretty crazy spikes in "OSD Read Latencies" and
"OSD Write Latencies" on the dashboard. Most of the time everything is
well but then for periods of time, 1-4 hours, latencies will go to 10+
seconds for one or more OSDs. This also happens outside scrub hours
and it is not the same OSDs every time. The OSDs affected are HDD with
DB/WAL on NVMe.
Log snippet:
"
...
2024-03-22T06:48:22.859+0000 7fb184b52700 1 heartbeat_map is_healthy
'OSD::osd_op_tp thread 0x7fb169898700' had timed out after 15.000000954s
2024-03-22T06:48:22.859+0000 7fb185b54700 1 heartbeat_map is_healthy
'OSD::osd_op_tp thread 0x7fb169898700' had timed out after 15.000000954s
2024-03-22T06:48:22.864+0000 7fb169898700 1 heartbeat_map
clear_timeout 'OSD::osd_op_tp thread 0x7fb169898700' had timed out
after 15.000000954s
2024-03-22T06:48:22.864+0000 7fb169898700 0 bluestore(/var/lib/ceph/
osd/ceph-112) log_latency slow operation observed for submit_transact,
latency = 17.716707230s
2024-03-22T06:48:22.880+0000 7fb1748ae700 0 bluestore(/var/lib/ceph/
osd/ceph-112) log_latency_fn slow operation observed for
_txc_committed_kv, latency = 17.732601166s, txc = 0x55a5bcda0f00
2024-03-22T06:48:38.077+0000 7fb184b52700 1 heartbeat_map is_healthy
'OSD::osd_op_tp thread 0x7fb169898700' had timed out after 15.000000954s
2024-03-22T06:48:38.077+0000 7fb184b52700 1 heartbeat_map is_healthy
'OSD::osd_op_tp thread 0x7fb169898700' had timed out after 15.000000954s
...
"
"
[root@dopey ~]# ceph -s
cluster:
id: 8ee2d228-ed21-4580-8bbf-0649f229e21d
health: HEALTH_WARN
1 failed cephadm daemon(s)
Low space hindering backfill (add storage if this doesn't
resolve itself): 1 pg backfill_toofull
services:
mon: 5 daemons, quorum lazy,jolly,happy,dopey,sleepy (age 3d)
mgr: jolly.tpgixt(active, since 10d), standbys: dopey.lxajvk,
lazy.xuhetq
mds: 1/1 daemons up, 2 standby
osd: 540 osds: 539 up (since 6m), 539 in (since 15h); 6250
remapped pgs
data:
volumes: 1/1 healthy
pools: 15 pools, 10849 pgs
objects: 546.35M objects, 1.1 PiB
usage: 1.9 PiB used, 2.3 PiB / 4.2 PiB avail
pgs: 1425479651/3163081036 objects misplaced (45.066%)
6224 active+remapped+backfill_wait
4516 active+clean
67 active+clean+scrubbing
25 active+remapped+backfilling
16 active+clean+scrubbing+deep
1 active+remapped+backfill_wait+backfill_toofull
io:
client: 117 MiB/s rd, 68 MiB/s wr, 274 op/s rd, 183 op/s wr
recovery: 438 MiB/s, 192 objects/s
"
Anyone know what the issue might be? Given that is happens on and off
with large periods of time in between with normal low latencies I
think it unlikely that it is just because the cluster is busy.
Also, how come there's only a small amount of PGs doing backfill when
we have such a large misplaced percentage? Can this be just from
backfill reservation logjam?
Mvh.
Torkil
--
Torkil Svensgaard
Systems Administrator
Danish Research Centre for Magnetic Resonance DRCMR, Section 714
Copenhagen University Hospital Amager and Hvidovre
Kettegaard Allé 30, 2650 Hvidovre, Denmark
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
--
Alexander E. Patrakov
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
--
Torkil Svensgaard
Systems Administrator
Danish Research Centre for Magnetic Resonance DRCMR, Section 714
Copenhagen University Hospital Amager and Hvidovre
Kettegaard Allé 30, 2650 Hvidovre, Denmark
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
