Hi,

after upgrading one of our production clusters from Ceph 19.2.3 to 20.2.1
we observed that deep scrubbing is no longer keeping up with the configured
schedule.

The cluster is healthy apart from:

HEALTH_WARN
400 pgs not deep-scrubbed in time

The number of overdue PGs is steadily increasing (e.g. 0 -> 400 within 4
days).

Cluster
Ceph 20.2.1 (upgraded from 19.2.3 approximately one week ago)
18 OSDs
NVMe SSDs
2208 PGs
~27 TiB user data
~64% raw utilization
No recovery/backfill
All PGs are active+clean
18 osds: 18 up, 18 in
flags: noautoscale

OSD latencies are essentially zero:

ceph osd perf
commit_latency = 0 ms
apply_latency = 0 ms
Scrub configuration

The scrub window has been unchanged for a long time:

osd_scrub_begin_hour = 21
osd_scrub_end_hour   = 6

The cluster kept up with this configuration before the upgrade.

The scheduler configuration is:

osd_op_queue = mclock_scheduler
osd_mclock_profile = high_recovery_ops

This profile was already configured before the upgrade.

Observation:

I collect the ceph_pg_scrubbing Prometheus metric.

Before the upgrade (19.2.3):

scrubbing starts around 23:00
continues throughout the whole scrub window
there are continuously 1-5 PGs being scrubbed during the night

After the upgrade (20.2.1):

there is a large initial burst (around 30 PGs scrubbing simultaneously)
after that, scrubbing almost completely stops
only occasional individual PGs are scrubbed during the rest of the scrub
window

As a result, the number of overdue deep scrubs increases every day.

Questions:
Were there any changes to scrub scheduling or mClock behavior between
19.2.3 and 20.2.1 that could explain this?
Is this a known issue in 20.2.1?
Is high_recovery_ops known to affect deep scrub scheduling differently in
Ceph 20 compared to Ceph 19?
Are there any debug commands or perf counters that would help determine why
new scrub jobs are not being scheduled after the initial burst?

I'd appreciate any suggestions on what to investigate next.

Thanks!

Kamil Madac
_______________________________________________
ceph-users mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to