pg_autoscale_mode: on, warn, and off

Why this matters in production

Choosing pg_num by hand was, for years, the single most common way to misconfigure a Ceph cluster. Pick it too low and every OSD carries a few enormous PGs, so balance is coarse and recovery is lumpy. Pick it too high and the OSDs spend memory and CPU on PG bookkeeping that buys nothing. The autoscaler exists to take that decision off your plate — but it is a manager module making an inference from pool statistics, and you still need to know what it will do before you let it do it.

The three modes

ceph osd pool get rbd-prod pg_autoscale_mode
# pg_autoscale_mode: on

off — the module ignores the pool entirely. No suggestions, no warnings, no changes. pg_num is whatever you last set.

warn — the module computes what it thinks pg_num should be, and if the current value is far enough off, raises a health warning. It never changes anything. This is the “tell me, don’t touch it” mode.

on — the module computes the target and applies it, adjusting pg_num in the background over time.

Set it per pool, or set the default for pools created from now on:

ceph osd pool set rbd-prod pg_autoscale_mode warn
ceph config set global osd_pool_default_pg_autoscale_mode warn

What “applies it” actually means

on does not slam pg_num to the new value. It moves pg_num toward the target, and each step causes PG splits or merges, which cause data movement. The pace is governed by the usual backfill throttles, so an autoscaler change looks like an ordinary rebalance in ceph -s.

The important consequence: an autoscaler decision made at 03:00 produces recovery traffic at 03:00. You did not schedule it and you were not watching.

Where the module lives

pg_autoscaler is a manager module, enabled by default:

ceph mgr module ls | grep pg_autoscaler

If the active manager is down or the module is disabled, no suggestions appear and no scaling happens — a silent failure mode worth knowing about when autoscale-status comes back empty.

Quiz

Production discipline

Record the mode for every pool in the same place you record its size and rule — a pool whose autoscale mode nobody remembers setting is a pool that will move data at an inconvenient hour. When you set warn, commit to actually reading the warnings; a permanently-ignored health warning trains the team to ignore the health output entirely.

Cross-course references

• Kubernetes: pool PG counts affect the latency of RBD volumes backing StatefulSet PVCs
• Linux: manager module scheduling is ordinary process work — the mgr is a userspace daemon like any other