How it works

You describe the machine you need and the command to run. rrun finds the machine, runs the command, streams the logs back and cleans up afterwards. There is nothing for you to host or keep running.

A job, start to finish

1. Pack. o41-rrun run packs the folder and uploads it, unless rrun already has that exact content. The job is queued.
2. Find a machine. If one of your reserved machines is free and fits, the job takes it, since it costs nothing; next comes an idle cloud machine that fits. Either starts the job in 2–3 seconds. Otherwise rrun asks every connected cloud account that takes this kind of job for offers and launches the cheapest one that fits (provisioning). A cloud without capacity falls through to the next offer. --reserved-only waits for a reserved machine instead, and --cloud-only never takes one. When nothing fits, rrun keeps trying until --queue runs out (30 minutes by default), then the job fails with no_capacity.
3. Boot. The new machine starts a small agent that connects back to rrun, downloads your folder into a fresh directory and runs the command (starting, then running). A cold machine takes 30–240 seconds from launch to running, depending on the cloud and the image.
4. Run. Logs and GPU, CPU and RAM telemetry stream back while the command runs. o41-rrun logs -f and the job's page in the web app show them live.
5. Finish. The command's exit code ends the job as succeeded, failed or cancelled.
6. Keep warm, then delete. A cloud machine waits for the job's --keep (5 minutes by default) in case your next job fits, then it is deleted together with its disks. A reserved machine simply waits for the next job.

If a machine never comes up, the job goes back to step 2 and tries another offer. If the machine fails while the command runs, the job also goes back to step 2, keeping its placement, and the command runs again, up to --restarts times (10 by default; see Restarts). A cancel at any point ends the job for good.

When things go wrong

• Lost machine. A cloud machine that stops reporting is deleted, and its job restarts on another machine. A reserved machine that stops reporting, or reboots, mid-job loses that job the same way, and is offline until its agent is back. With no restarts left, the job fails with machine_lost, or preempted when the cloud took a spot machine back.
• Timeout. --timeout stops the command and fails the job with timeout.
• Cancel. The command gets SIGTERM, then SIGKILL after 30 seconds.
• Leftovers. rrun regularly checks each cloud account for machines and disks named for your organization and that account that no job owns, and deletes them. See Machines and costs.

Logs

Logs and telemetry are kept in rrun's storage for 30 days, or in your own bucket as long as you like, and stay readable after the machine is gone.

Security

• Cloud credentials are encrypted and only used to talk to that cloud. Google Cloud can connect without a key at all; see Clouds.
• Each machine gets its own token, good only for its own job, and the token ends with the machine. A reserved machine's key is bound to that machine and works until the machine is removed; rrun keeps only its hash.
• An organization API key can manage jobs only. See API keys and agents.