The replicated rule — the form you will write most often

Why this matters in production

Nearly every pool on nearly every cluster uses a rule of this shape. Writing them correctly and verifying them is a small, frequently used skill.

The form

rule rack_replicated {
  id 1
  type replicated
  step take default
  step chooseleaf firstn 0 type rack
  step emit
}

Creating rules with the CLI

Hand-editing the map is unnecessary for standard rules:

NAME=acme
ROOT=root
FAILURE_DOMAIN=vm-db-01
CLASS=class
ceph osd crush rule create-replicated ${NAME} ${ROOT} ${FAILURE_DOMAIN} [${CLASS}]

ceph osd crush rule create-replicated rack_rule default rack
ceph osd crush rule create-replicated nvme_host default host nvme
ceph osd crush rule ls
ceph osd crush rule dump rack_rule

The CLI assigns the id, writes correct syntax, and cannot renumber existing rules — which is the main hazard of hand-editing.

Common variants

# host-level, the default
ceph osd crush rule create-replicated host_rule default host

# rack-level
ceph osd crush rule create-replicated rack_rule default rack

# NVMe only, host separation
ceph osd crush rule create-replicated nvme_rule default host nvme

# restricted to one rack (isolation, not redundancy)
ceph osd crush rule create-replicated racka_only rack-a host

The last is the one to use sparingly — it caps the pool at that subtree’s capacity permanently.

Verifying the outcome

PGID=12.1a
OSD=12
ceph osd pool get rbd-vms crush_rule
ceph pg ls-by-pool rbd-vms | head -5
ceph pg map ${PGID}
ceph osd find ${OSD} | jq '.crush_location'

Sample a few PGs and confirm the acting sets span the intended domains. Reading the rule shows the intent; reading the acting sets shows the result.

Quiz

Production discipline

Create rules with ceph osd crush rule create-replicated rather than by hand-editing the map — the CLI writes correct syntax and cannot renumber existing rules, which is the main hazard of editing. Always validate with crushtool --test between creating and assigning, since that is the window in which being wrong costs nothing. And use firstn 0 in every replicated rule, because a hard-coded count silently undersizes every PG the day someone raises size.

Cross-course references

• Ceph: Part XIII (CRUSH Fundamentals) for what the steps do.
• Ceph: Part XVI (Device Classes) for class-aware rules.
• Ceph: Part XVII (Pools) for assigning rules to pools.