The maintenance reserve is a scheduling decision

Why this matters in production

The maintenance reserve is the only line in a capacity plan whose size is set by a policy rather than by measurement, and it costs real hardware.

Maintenance costs no capacity until it does

ceph -s
ceph orch host ls
ceph osd set-group noout ceph-04
ceph orch host maintenance enter ceph-04

With the flag set, the absent OSDs stay in the map, no PG is remapped and no byte moves. The cluster is short of redundancy, not short of space. The capacity cost appears only in the case where the host does not come back — and at that point it is exactly the failure reserve, not a separate one.

Two policies, two reserves

At six hosts that is the difference between 17 and 33 per cent of usable capacity. At twenty hosts it is 5 against 10, which is why large clusters adopt the relaxed policy and small ones cannot afford to.

The chance that maintenance becomes a failure

ceph orch host maintenance exit ceph-04
ceph -s
ceph osd unset-group noout ceph-04

You cannot buy a fraction of a host, so the probability does not size the reserve. What it decides is the policy: at small N accept the strict policy and the patch latency it causes; at larger N simply hold the second host of reserve and stop scheduling around it.

Putting the policy where the plan can see it

cat >> /secure/ceph-capacity-plan.txt <<'EOF'
maintenance policy:   one host at a time, only from HEALTH_OK
maintenance reserve:  0 TiB beyond the failure reserve
accepted consequence: patching blocked while backfilling; log each deferral
revisit at:           12 hosts, or after 3 deferrals in a quarter
EOF

Quiz

Production discipline

Record the maintenance policy next to the maintenance reserve, because one determines the other and a plan showing only the number cannot be reviewed. Log every deferred maintenance item — a pattern of deferrals is the evidence that the strict policy has stopped being affordable.

Cross-course references

• Kubernetes: a PodDisruptionBudget is the same trade written as a scheduling rule
• Linux: patch cadence and redundancy budget are the same decision in two documents