Retries and Recovery

Retry Model

Retries are controlled by RestartPolicy.

Core behavior:

• retries stop at max_retries
• countdown is linear or exponential based on exponential_backoff
• optional jitter randomizes retry timing
• optional exception allowlist narrows retry scope

Profiles

CeleryLanguage defines profile presets:

• none
• standard
• aggressive

Profiles are merged first, explicit constructor fields override profile defaults.

Policy Sources

Policy can come from:

1. language defaults
2. profile defaults
3. explicit constructor fields
4. per-dispatch policy passed to say(...)

Recovery Capabilities

• Journal.restart_recent(...) for restarting recent failed runs.
• Language.resubmit(...) for replaying a recorded run from stored payload.
• FastAPI example includes admin endpoints for full-job requeue and compiled-node requeue.

Event Signals to Watch

Useful events for operations and diagnostics:

• job.retry_scheduled
• subjob.retry
• job.failed
• workflow.lowering_failed