test-a-data-pipeline-backfill

# test-a-data-pipeline-backfill Backfill from a fixed input contract and prove exactly which outputs it may change. ## When to use - Use for bug repair, schema change, new logic, late data, model rebuild, or historical restoration. - Do not run against production until mutation scope and downstream consumers are authorized. ## Procedure 1. Define reason, datasets, key or partition range, input snapshot, logic version, outputs, downstream consumers, and acceptance thresholds. 2. Reproduce the defect and expected correction on a representative bounded fixture. 3. Make processing idempotent with stable record identity, versioning, deduplication, and safe retry. 4. Test late, duplicate, missing, deleted, out-of-order, schema-old, and boundary records. 5. Compare before and after row counts, keys, values, aggregates, lineage, quality, and representative consumers. 6. Load-test read, compute, write, queue, warehouse, source, and downstream capacity with throttling and stop controls. 7. Rehearse interruption, resume, partial failure, bad output quarantine, and rollback or forward correction. 8. Canary a small production range, reconcile, expand in waves, and preserve manifests. ## Done - A backfill test plan and run manifest record scope, snapshot, logic, identity, fixtures, capacity, waves, outputs, and owners - Idempotency, boundary, interruption, retry, duplicate, schema, before-after, downstream, capacity, and rollback checks verify readiness