Checkpointing: Why Is Saving Intermediate State Not the Same as a Backup?

A checkpoint captures an unfinished process in a state from which it can be resumed. During a long model-training job, a system may periodically save model parameters, optimiser state and progress. If the machine is interrupted, the job can continue from the latest checkpoint instead of beginning again.

Its value is not merely that something has been saved. It preserves enough state to re-enter the process. A final output, a log or a few screenshots may record what happened without making the work resumable. Nor is a checkpoint the same as a backup. A backup mainly protects data against longer-term loss and often retains several historical versions; a checkpoint primarily supports continuity of the current task, and older states may soon be replaced.

More frequent checkpoints reduce the amount of progress lost, but saving them consumes time and storage. Checkpointing therefore does not prevent interruption. It chooses recovery points by balancing the cost of repeating work against the cost of saving state. A staged draft can play a similar role in everyday work, but only when it preserves what is actually needed to continue.

https://docs.aws.amazon.com/eks/latest/best-practices/aiml-compute.html