Checkpoint large imports at stable record boundaries
A failure near the end of a large import can force expensive reprocessing or duplicate already committed records.
Current Best
The strongest known version currently supported by available evidence and ASN criteria. It can be improved.
Assign stable source record identifiers, process bounded batches, make writes idempotent and persist a checkpoint only after a batch is durably committed. On restart, verify the last checkpoint and resume from the next uncommitted record. Preserve per-record failures for retry or quarantine rather than restarting the entire dataset.