Recommendations

Cetacean watches your cluster and lists what looks wrong at /recommendations in the dashboard—services with no health check, replicas that keep dying, containers sized far above what they use. Findings refresh on their own; you never trigger a scan.

It’s on by default. Turn it off with server.recommendations.

What gets flagged

Each finding names a service, a node or the cluster, and carries a severity of info, warning or critical.

no-healthcheck

The service defines no health check, so Swarm can’t tell a hung container from a working one

no-restart-policy

Restart policy is none, so a failed task is never replaced

single-replica

One replica, so any node problem is an outage

manager-has-workloads

A manager node is active, so Swarm may schedule ordinary tasks on it alongside cluster management. Not raised unless a ready, active worker could take those tasks

uneven-distribution

The busiest node runs more than three times the tasks of the quietest

flaky-service

More than 5 involuntary task failures within the lookback window

node-disk-full

Node disk above 90%

node-memory-pressure

Node memory above 90%

over-provisioned

Reserved far more CPU or memory than it uses

approaching-limit

Usage climbing towards its limit

at-limit

Usage at its limit, so it’s being throttled or is at risk of being killed

no-limits, no-reservations

No limit or reservation set, so Swarm can’t schedule it well

Act on a finding

Findings that have an obvious fix get an Apply suggested value button:

All three also need write permission on the service—the button only appears for callers who have it. Everything else is reported for you to act on yourself.

Without Prometheus

Cetacean still flags configuration and topology problems, and still catches flaky services, using Docker state alone. What disappears is everything measured: the whole sizing group, plus node-disk-full and node-memory-pressure. Nothing errors—you simply see fewer findings.

Sizing needs cAdvisor and the node checks need node-exporter. See Monitoring for the setup.

Tune the sizing thresholds

Sizing is the only group you can tune, through sizing.thresholds.over_provisioned, sizing.thresholds.approaching_limit and sizing.thresholds.at_limit—the fractions of the reservation or limit at which each finding appears. sizing.headroom_multiplier sets how much room the suggested value leaves above observed usage, and sizing.thresholds.lookback how far back Cetacean looks, which is also the window flaky-service counts failures in.

API

GET /recommendations returns the current findings and a severity summary; the MCP get_recommendations tool serves the same data. See the API reference for the response schema.