Checkpoints and resuming

How workspaces, FairGPU cloud checkpoints and downloads protect your progress, what they cost, and how to resume on the same machine or any other.

Consumer PCs reboot, hosts go offline, and sometimes a host needs the PC back. FairGPU gives you three layers so a lost rental costs minutes, not days.

Workspaces (same host)

Tick Keep a workspace on this machine in the rent wizard. It mounts at /workspace, keeps your files between sessions on that host, and shows under Workspaces with Resume, Grow and Delete. It is on the host's PC: a host that reinstalls Docker loses it (you are told, and the last storage tick is refunded), so pair it with a checkpoint or sync target for anything precious.

Checkpoints (any host)

A checkpoint is a compressed archive of /workspace uploaded by the host's agent to FairGPU storage through a one-time link. The host never holds storage credentials; the bucket provider encrypts in transit and at rest.

Three ways one is made:

1. Automatically: pick every 15 / 30 / 60 min under Progress safety when renting (only while the rental runs and is not idle; one copy per rental is kept, the latest).
2. On a reclaim notice: the moment a host gives notice, an interrupt checkpoint is taken, and another best-effort one before the container is removed.
3. By hand: Progress safety → Save checkpoint now on the rental page, or fairgpu snapshot save "before eval" inside the container.

You see the progress on the card and the timeline (snapshot.started → progress → done). Every checkpoint is listed under Workspaces → Checkpoints with its size, monthly cost and expiry (30 days by default; you get a notice 72 hours before).

Limits: 5 GB per checkpoint in this version; keep huge datasets re-downloadable rather than checkpointed.

Cost example

A 3 GB workspace at the platform price of $0.04 per GB-month with 2 GB free: 1 GB billable ≈ $0.04 per month, billed per second while the checkpoint exists. Delete it and billing stops.

Resume

Same machine

An INTERRUPTED rental (host went away) shows Resume: a new rental on the same machine, same price, same workspace, for the minutes you had left. If the host is offline, leave an auto-resume order and it starts when the machine is back, within the 24-hour window.

Another machine

With a READY checkpoint, an interrupted or stopped rental shows Resume on another machine: pick any online machine (those with the image cached are listed first); the checkpoint is restored into the new /workspace before the container starts, so your pip packages and files are there from the first second. The stepper shows Restoring with progress. The same button exists on each checkpoint under Workspaces, and the marketplace link carries the checkpoint into the rent wizard if you want to tweak settings first.

Restore into a fresh rental

Any checkpoint, including one you uploaded from your computer (.tar.zst, .tar, .zip up to 5 GB), can seed a rental: Restore next to it, or Rent with this flow from the marketplace.

Download

Progress safety → Download workspace streams the latest checkpoint as .tar.zst (a fresh one is created first if the last is older than 10 minutes). Unpacking commands per OS are in Move files. For live transfers of individual files use scp / rsync / SFTP.

Honest notes

• Checkpoints are stored on FairGPU's cloud storage (Cloudflare R2), encrypted in transit and at rest by the provider; the host PC only gets a one-time upload link.
• Servers without storage configured show unavailable on this server where these buttons would be; workspaces and your own sync targets still work.
• Retention: 30 days, extended as long as your balance covers it; after a LOW_BALANCE notice you get 72 hours before deletion.