ceph-aiops

Ceph AIops

Disclaimer: Community-maintained open-source project, not affiliated with, endorsed by, or sponsored by the Ceph project or any storage vendor. Product and trademark names belong to their owners. Source at github.com/AIops-tools/Ceph-AIops under the MIT license.

Governed Ceph operations via the ceph-mgr Dashboard REST API — 37 MCP tools, every one wrapped with the bundled @governed_tool harness: a local unified audit log under ~/.ceph-aiops/, token/runaway budget guard, undo-token recording, and descriptive risk tiers. The Dashboard password is stored encrypted (~/.ceph-aiops/secrets.enc, Fernet + scrypt) — never plaintext on disk. The flagship cluster_health turns raw HEALTH_WARN/ERR check codes into plain-language cause + suggested action.

Standalone: the governance harness is bundled in the package (ceph_aiops.governance) — ceph-aiops has no external skill-family dependency. Works against vanilla ceph-mgr (cephadm / hypervisor-bundled / MicroCeph); no croit, no Kubernetes.

What This Skill Does

Totals: 37 tools — 17 read, 18 write, 2 undo. The MCP server exposes all 37; the CLI is a convenience subset.

Quick Install

uv tool install ceph-aiops
ceph-aiops init       # interactive wizard: mgr host/port/username + encrypted Dashboard password
ceph-aiops doctor

When to Use This Skill

• Decode a HEALTH_WARN/ERR state (cluster_health / health detail) — cause + action per active check (PG_DEGRADED, OSD_NEARFULL, SLOW_OPS, MON_DOWN, LARGE_OMAP_OBJECTS, …)
• One-shot triage (overview): HEALTH status + active checks + OSD up/in counts
• Inspect OSDs (osd_tree / osd_df most-full first / osd_perf slowest first), PGs (pg_summary / pg_dump_stuck / scrub_status), pools (pool_ls / pool_df usable capacity)
• Investigate slow requests (slow_ops), MDS trimming lag (cephfs_status), RGW large-omap (rgw_status), mon quorum (mon_status), and days-to-nearfull (capacity_forecast)
• Safely drain + purge an OSD, change pool size/quota, or throttle a slow rebalance (governed writes with dry-run + undo)

Do NOT use when the target is not Ceph (a hypervisor, another storage appliance, a backup product, a container cluster, or a network device). Route those to the appropriate other AIops-tools skill.

Related Skills — Skill Routing

Common Workflows

1. "The cluster went HEALTH_WARN overnight" — decode it (read-only)

1. ceph-aiops doctor → confirm the mgr Dashboard is reachable and the JWT login works before trusting anything else
2. ceph-aiops overview → HEALTH status, the list of active check codes, and OSD up/in counts in one shot
3. ceph-aiops health detail (MCP: cluster_health) → each active check translated into what it means, the likely cause, and a suggested action
4. Drill into the implicated resource: PG_DEGRADED → pg_dump_stuck; OSD_NEARFULL → ceph-aiops osd df (most-full first); SLOW_OPS → slow_ops; MON_DOWN → mon_status; LARGE_OMAP_OBJECTS → rgw_status
5. Failure branch: if doctor fails on auth, the Dashboard password is wrong or the store is locked — re-run ceph-aiops secret set (or export CEPH_AIOPS_MASTER_PASSWORD for non-interactive use). If doctor fails on reachability, the mgr dashboard module is likely not enabled; no read is issued against an unauthenticated session.

2. Retire a failing OSD: drain, mark out, purge (governed)

1. ceph-aiops health detail → confirm the OSD is genuinely the problem (e.g. OSD_SLOW_PING_TIME, repeated SLOW_OPS on one id) rather than a cluster-wide symptom
2. ceph-aiops osd df → confirm the id, and that the remaining OSDs have room to absorb its data before you drain anything
3. ceph-aiops osd reweight 0.0 → start a gradual drain; reversible, the prior CRUSH weight is captured as the undo descriptor
4. ceph-aiops osd out --dry-run, then re-run without --dry-run → high risk, double confirmation, needs CEPH_AUDIT_APPROVED_BY
5. Wait for ceph-aiops health status / pg_summary to show all PGs active+clean — do not purge while backfill is running
6. ceph-aiops osd purge --dry-run, then re-run without --dry-run → high, irreversible
7. Failure branch: if client I/O tanks during the drain, stop and reverse — ceph-aiops undo list then ceph-aiops undo apply restores the prior weight (and osd_mark_in reverses the mark-out). Purge has no undo, which is exactly why it comes last and after active+clean.

3. Recovery is starving client I/O

1. ceph-aiops health detail → confirm the cluster is actually backfilling/recovering (PG_DEGRADED, PG_BACKFILL_FULL) rather than hitting a different bottleneck
2. slow_ops → check whether client requests are genuinely being blocked, and by which OSDs
3. throttle_recovery(max_backfills=1, recovery_max_active=1) → med risk, reversible; the prior osd_max_backfills / osd_recovery_max_active are captured as the undo descriptor
4. Re-check slow_ops and pg_summary — recovery is slower but client latency should recover
5. Once the cluster is quiet, raise the values back (or ceph-aiops undo apply to restore the exact prior settings)
6. Failure branch: if throttling does not help, the bottleneck is not recovery — go back to osd_perf (slowest OSDs first) and mon_status; do not keep lowering the throttle, you will only extend the degraded window.

4. A pool is running out of usable capacity

1. ceph-aiops overview → look for POOL_NEARFULL / OSD_NEARFULL among the active checks
2. pool_df → per-pool usage with usable capacity = raw ÷ size (a size=3 pool reports a third of raw — this is where most "but the disks aren't full" confusion comes from)
3. capacity_forecast → days-to-nearfull at the current fill rate, so you know whether this is a this-week problem or a this-quarter one
4. Buy time reversibly first: set_pool_quota (med, undo → prior quota) or set_pool_autoscale (med, undo) to let pg_num track the new size
5. Only if a replica change is genuinely the answer: set_pool_size --dry-run then the real call — high risk, because lowering size reduces durability and any change forces cluster-wide data movement
6. Failure branch: if the resulting rebalance saturates the cluster, apply workflow 3 (throttle_recovery) rather than reverting the size mid-flight; if the size change itself was wrong, ceph-aiops undo apply replays the recorded prior value — but expect a second full rebalance.

Governance & Safety

The skill delivers reads and writes and records them; it does not decide whether a write is permitted. That is your agent's judgement, or the permission of the account you connect it with (a ceph-mgr Dashboard account with a read-only role — writes then fail at the mgr). There is no read-only switch, policy file, or approval gate.

• Audit is the guarantee, and it is not bypassable. Every operation — MCP and CLI alike — is logged to ~/.ceph-aiops/audit.db (relocatable via CEPH_AIOPS_HOME): params, result, status, duration, and the risk tier. The CLI writes the same row the MCP path does.
• CEPH_AUDIT_APPROVED_BY / CEPH_AUDIT_RATIONALE are optional annotations recorded on the audit row (who/why); they are never required and never block.
• Runaway guard — a safety backstop, not authorization: the same call looped in a tight window trips a circuit breaker. Disable with CEPH_RUNAWAY_MAX=0.
• Destructive writes support --dry-run / dry_run=True and double confirmation at the CLI.
• Reversible writes fetch the real before-state and record an inverse descriptor (osd_reweight→restore prior weight, cluster_flag_set→toggle back); irreversible ops (osd_purge, pool_delete, RBD deletes) record only the before-state.

References

• references/capabilities.md — full tool → API-path → returns reference
• references/cli-reference.md — CLI command reference
• references/setup-guide.md — onboarding, credentials, and connectivity