A Run Job runs one command in a container on temporary GPU or CPU compute. The command runs to completion, its output streams back to you, and the compute is released the moment it exits, so there's nothing to stop afterwards.
Run Jobs suit work with a clear end, such as:
For interactive work, such as notebooks, debugging or a model server that stays up, use an instance.
A Run Job runs your command in one of these:
Run Jobs bill per second for compute and container disk, from the moment the container starts until the command exits. Time spent waiting for compute and downloading the image isn't billed.
Before a Run Job starts, credit for its full time limit is set aside. When it ends, you're charged only for the time it ran, and the rest is released. A job whose time limit your credit can't cover waits until you add credit, so set the time limit close to what the job needs. See Billing.
Run Jobs can be interrupted to make room for other work. If you allow retries, an interrupted job starts again from the beginning; otherwise it stops there. Write checkpoints to a network volume so a retry can resume where it left off.