| name | ingest-backfill-runbook |
| description | Plan and execute safe ingestion replay or backfill runs with partition locks, idempotency checks, evidence outputs, and rollback notes. |
| version | 0.1.0 |
| owner | data-platform |
| status | active |
| tags | ["ingestion","backfill","replay","dagster"] |
| not_for | ["Ad hoc SQL updates without source manifest or partition scope.","Destructive data deletion requests."] |
| inputs | ["data/seed_manifests/*.json","pipelines/data_factory/*"] |
| outputs | ["reports/week09/backfill_runbook.md"] |
| requires | ["data-contract-lint"] |
| compatible_agents | ["codex","claude-code","mcp-compatible-agent"] |
| artifacts | ["reports/week09/backfill_runbook.md"] |
| evals | ["tests/integration/test_week06_backfill_plan.py"] |
Ingest Backfill Runbook
Use this skill when a task asks for replay, backfill, idempotent ingestion,
partition repair, or recovery after a failed ingest.
Procedure
- Identify the affected data release, source manifest, partitions, and
downstream assets.
- Confirm idempotency keys and partition boundaries before running anything.
- Produce a dry-run plan with expected row counts and affected artifacts.
- Execute only bounded runs. Record command, operator, timestamps, and output.
- Write recovery evidence under
reports/week09/ and include rollback notes.
Safety Boundaries
- Never run broad unscoped backfills.
- Never overwrite prior reports without preserving the original evidence.
- Escalate if partition locks or source fingerprints are missing.