Snapshotting Instances Without Powering Off | Generated by AI

Home 2026.05

Question: For AMD Dev Cloud and DigitalOcean — can you create snapshots of a running instance without powering it off first?

Answer:

These two platforms have different behaviors:

DigitalOcean

No — you need to power off first for a consistent snapshot.

DigitalOcean does support “live snapshots” (snapshot while running), but they explicitly warn that this can result in inconsistent/corrupted filesystem state — essentially a crash-consistent snapshot, not application-consistent. For anything stateful (databases, in-progress training runs with open files), powering off is the safe path.

Their docs recommend:

# Power off cleanly before snapshotting
doctl compute droplet-action power-off <droplet-id> --wait
doctl compute droplet-action snapshot <droplet-id> --snapshot-name "my-snapshot"

Live snapshots on DO are essentially “best effort” — fine for stateless machines, risky for active workloads.

AMD Dev Cloud

The snapshot page at devcloud.amd.com/images/snapshots is specific to their environment. AMD Dev Cloud instances (especially MI300X nodes billed at ~$2/hr) are typically ephemeral compute nodes where:

I’d check the actual UI behavior — if the “Create Snapshot” button is greyed out while running, that’s your answer. If it’s active, it’s likely crash-consistent at best.

Bottom Line

Platform Live Snapshot? Safe for ML workloads?
DigitalOcean Technically yes, but risky Power off first
AMD Dev Cloud UI-dependent, likely needs stop Stop instance first

Practical advice for your use case: Before snapshotting an AMD Dev Cloud instance mid-training-run, checkpoint your model weights explicitly (they should already be on disk), then stop the instance cleanly before snapshotting. Snapshots are for environment/dependency preservation — not checkpoint state.


Back Donate