17. Retrying failed jobs keeps their completed steps
Date: 2026-10-05
Status
Accepted
Context
A FAILED job could only be run again with SQL, e.g. after fixing the bug that made it fail.
Decision
Bikesched.retry(jobId) puts a FAILED job back to SCHEDULED, due right away, with its
attempts reset and its compensation and cancellation state cleared, in one statement recorded in
the job history (“retried”). Its steps are kept: completed ones return their stored results,
compensated ones run again. last_error is kept until the next failure. Retrying an occurrence of
a recurring job throws an IllegalArgumentException for now, rather than returning false, so
the caller sees why: its next occurrence was already scheduled when it failed, and the partial
unique index allows only one pending occurrence per recurring job.
Consequences
- Failed jobs continue where they failed instead of starting over.
- Only possible until the job retention deletes the job;
COMPENSATION_FAILEDjobs and failed occurrences of recurring jobs cannot be retried this way.