The scrub suppression flags

Why this matters in production

Scrub suppression is the most commonly left-set flag because nothing visible happens while it is on.

The two flags

ceph osd set noscrub
ceph osd set nodeep-scrub
ceph osd dump | grep flags

Neither stops scrubs already in progress.
A deep scrub of a large PG continues after the flag is set.

ceph pg dump pgs | grep -c scrubbing

Waiting for the in-progress scrubs to drain is what makes the suppression effective, and measuring immediately after setting the flag shows no change.

Resumption behaviour

While suppressed, last_scrub_stamp does not advance while time does.
Every PG's overdue margin grows together.
On resumption, all become eligible simultaneously.
Many exceed max_interval, which forces them past the load and time checks.

ceph pg dump pgs 2>/dev/null | awk 'NR>1 {print $23}' | sort | head -3
ceph health detail | grep NOT_DEEP_SCRUBBED

# resume gradually after a long suppression
ceph config set osd osd_max_scrubs 1
ceph osd unset nodeep-scrub

Lowering the concurrency before resuming spreads the burst.

The coverage cost

Deep scrub is the only verification of data that is not read.
Suppressing it for a month means a month of unverified cold data.
Corruption arising during that period is undetected until it resumes.

# how overdue is the oldest?
ceph pg dump pgs 2>/dev/null | awk 'NR>1 {print $23}' | sort | head -1

Per-pool alternatives

ceph osd pool set scratch noscrub true
ceph osd pool set scratch nodeep-scrub true
ceph osd pool get scratch nodeep-scrub

A per-pool flag expresses the actual intent — this pool does not need
scrubbing — without removing the check from every other pool.

# and the tuning alternatives, which reduce load without removing coverage
ceph config set osd osd_scrub_sleep 0.1
ceph config set osd osd_deep_scrub_interval 1209600
ceph config set osd osd_scrub_begin_hour 22
ceph config set osd osd_scrub_end_hour 6

Quiz

Production discipline

Extend the deep scrub interval and confine scrubs to a window rather than suppressing them — the load reduction is comparable and the coverage cost is bounded and measurable. Lower osd_max_scrubs before clearing a long-standing suppression, since every PG becomes eligible at once.

Cross-course references

• Kubernetes: disabling a periodic check because it is noisy removes what it verified
• Linux: deferring integrity checks indefinitely is functionally disabling them