delete_job

Hide a job from listings, keeping it for billing. Cost class: fast.

A soft delete. The row survives, so units keep answering and GET /v1/usage still counts the work; what stops is GET /v1/jobs returning it. Reading it by id still works, with deleted_at set — the same reasoning as an expired payload, where a 404 would conflate “gone” with “never existed”.

A job still queued or running is cancelled first, so hitting delete stops the compute rather than paying for a run nobody will ever see. A job that finished under us is simply hidden: there was nothing left to stop, and refusing the delete over a race the caller cannot see would be noise.

Temporal is not declared as a dependency here because asking for the client connects: most deletes are of finished runs and need it for nothing, so an outage must not fail work that is entirely in Postgres.

There is no project-level route to pair with ?project_id=: a run’s fan-out is one job per file, and the console deletes the group the same way it renames it — per job, over the ids it already holds.