The manager daemon — what it does and what it does not

Why this matters in production

The manager is the daemon most likely to be misread during an incident, because its failure produces alarming symptoms — no status, no dashboard, no metrics — while affecting no client at all.

What the manager does

• Collects and exposes metrics. OSD, pool, and PG statistics, and the Prometheus endpoint.
• Runs modules. Dashboard, balancer, orchestrator, telemetry, alerts, device health, and more.
• Serves the rich parts of ceph status. Capacity, I/O rates, and PG summaries come from the manager.
• Provides orchestration. ceph orch is the orchestrator module, which is a manager module.

What it does not do

• It holds no cluster map authority; monitors do.
• It is not consulted for any read or write.
• It performs no recovery and no placement.
• It stores no data.

flowchart LR
  M[Monitors: authoritative map] --> G[Manager: reads state, runs modules]
  G --> D[Dashboard]
  G --> P[Prometheus endpoint]
  G --> O[Orchestrator → cephadm]
  C[Clients] -.->|"never talks to MGR"| G
  C --> OS[OSDs]

Active and standby

Managers run active/standby. One is active; the others hold modules loaded but idle and take over if the active fails.

ceph mgr stat
ceph mgr dump | jq '.active_name, .standbys[].name'
ceph orch apply mgr --placement="2 ceph-01 ceph-02"
ceph mgr fail                     # force failover to a standby

Failover takes seconds. In-flight dashboard sessions reconnect; Prometheus scrapes may miss a point. Nothing else notices.

Two managers is the standard. More than two adds little, since only one is active and the second already provides failover.

Placement

Managers are lightweight on small clusters and can sit alongside monitors. On large clusters they deserve their own resources.

Place the two managers in different failure domains, for the same reason as monitors: a rack failure that takes both leaves the cluster blind at precisely the moment visibility matters most.

ceph orch ps --daemon-type mgr

Check that the two are not on the same host — the same silent drift that affects monitors affects managers.

Quiz

Production discipline

Run two managers in different failure domains and verify they are not on the same host, since the same silent placement drift that affects monitors affects managers. Alert on manager availability from outside the manager’s own metrics path, because otherwise its failure silences the alert that would report it. And on large clusters, check scrape duration against scrape interval before treating a dashboard gap as evidence about the cluster.

Cross-course references

• Ceph: Part LXXV (Prometheus Metrics) for the endpoint in detail.
• Ceph: Part LXXVIII (Monitoring the Monitoring Path) for the circularity problem.
• Ceph: Part X lesson on failover for the active/standby mechanics.