Five-minute quickstart

The examples below use Sparkrun, the current reference manager. It derives ColdSnap requests from placement and the recipe’s top-level coldsnap: block.

1. Verify your inputs

Use a digest-pinned runtime image, a pinned model revision, and a cluster whose hosts report a mutually supported snapshot driver. Confirm the recipe runs normally before capturing it.

2. Capture the initialized workload

sparkrun coldsnap capture recipe.yaml --cluster two-node

Capture starts the real engine, validates its response, and packages the admitted runtime state. It can take longer than a normal start because capsule construction happens once for subsequent restores.

3. Run from the captured state

sparkrun run recipe.yaml --cluster two-node

sparkrun run selects the compatible ColdSnap execution strategy, prepares and verifies the required artifacts, and replaces the serving workload only after preparation succeeds.

4. Check the response

Use the same validation prompt and serving endpoint you used before capture. A successful process restore is not sufficient: the workload must return the expected non-empty inference response.

Common next operations

# Publish capsules explicitly
sparkrun coldsnap publish recipe.yaml --cluster two-node

# Restore without the general run flow
sparkrun coldsnap restore recipe.yaml --cluster two-node

# Inspect the placed workload lifecycle
sparkrun coldsnap status recipe.yaml --cluster two-node