| name | job-observability |
| description | Observe automated jobs through trigger, logical operation identity, queue/wait time, attempts, execution, side effects, checkpoints, failures, completion, and stale/stuck detection. |
Job Observability
Use when scheduled, queued, event-triggered, or background automation needs to be diagnosable and operable at scale.
Procedure
- Define the logical job/operation identity separately from individual attempts so retries, node relocation, and replay can be correlated to one intent.
- Record trigger/source, tenant/resource scope, workflow/job type, revision/config version, enqueue/start/end time, executing node/worker when relevant, attempt count, and final state.
- Measure queue/wait age, runtime, completion/failure rate, retry rate, backlog/depth, concurrency, and resource/dependency pressure that can explain delayed work.
- Emit structured lifecycle events for accepted, claimed/started, checkpointed, retried, blocked/waiting, succeeded, failed, cancelled, expired, and reconciled states that exist in the automation.
- Correlate external side effects using safe provider/request IDs without logging credentials or unnecessary sensitive payloads.
- Detect stale/stuck work using explicit heartbeats, leases, deadlines, or progress checkpoints appropriate to the job model rather than one universal timeout.
- Distinguish retry churn from healthy throughput. A job eventually succeeding after many attempts may still indicate a degraded dependency or broken policy.
- Alert on actionable conditions: missed schedules/objectives, old queue age, repeated failure, retry exhaustion, stalled checkpoints, or systematic duplicate suppression/reconciliation anomalies.
- Provide a queryable history that supports replay/recovery decisions and can answer what ran, what changed, which side effects occurred, and why the job ended in its current state.
- Test telemetry during crash/restart, duplicate trigger, retry, stale worker, relocation to another Fleet node, partial side effect, and final reconciliation.
Decision rules
- Attempt logs alone are not job history; preserve the logical operation across attempts.
- Metrics show population health while event/run records explain individual failures; use both where the scale warrants it.
- Do not mark a job successful merely because its process exited zero if its intended side effect or durable outcome is unverified.
- Fleet may own node/task placement, but automation telemetry should carry node/attempt identity so relocation and correlated node failures are diagnosable.
Quality gate
Job observability is sufficient when operators can follow one logical job across triggers, attempts, nodes, checkpoints, and side effects; population-level backlog/failure health is visible; stuck work is detectable; alerts are actionable; and replay/recovery decisions can be made from recorded evidence.
