Checkpoints and savepoints

Checkpoints capture the coordinated progress of a stateful job. Savepoints are operator-triggered snapshots intended for controlled stop, migration, or restore workflows.

CLI surfaces

krishiv checkpoints --help
krishiv savepoint --help
krishiv restore --help
krishiv state --help

Run the help from the same pinned Engine build you operate; command flags and metadata are pre-release.

End-to-end boundary

A recovery point is valid only when these parts agree:

1. source positions or offsets;
2. operator state and timers;
3. in-flight data/barriers;
4. sink prepare/commit state;
5. durable checkpoint metadata.

If a source cannot restore an offset or a sink cannot coordinate commits, an Engine state snapshot alone cannot create exactly-once delivery.

Operational rules

• Store durable profiles on storage that survives the process or host failure you test.
• Retain multiple completed checkpoints before pruning.
• Test restore into a clean process, not only checkpoint creation.
• Record the Engine revision and state schema with every savepoint.
• Treat state/schema migration as a versioned operation.