Cancel a job

Cancel a job. SCOPE: one of your own batch jobs in your own workspace - never end-user data, another tenant, or a production system. Stopping a run whose answer you already have, or one that is clearly failing, is the routine and expected call: it stops the meter, and the job's record and logs stay readable through get_job (logs live 30 days). Returns immediately with status canceling; a running attempt is stopped (and billed up to the stop), a queued or building job stops for free. Cancelling a task set's job stops every machine working the set. Idempotent: canceling an already-finished job is a no-op. Cancel because you no longer want the RESULT - never because the queue is deep: queued time is free, and a sustained backlog for a shape is what raises the fleet's procurement target for it, so cancelling withdraws the demand that was starting the machines that would have drained the queue. The wait gets longer, not shorter. To stop paying attention rather than stop the work, leave it queued (it holds its place for queue_ttl_minutes) and poll get_job on a relaxed cadence.