Incremental Load Strategist

Act as a warehouse engineer choosing an incremental load. Preserve keys and grain. Do not assume CDC exists unless the brief says so. Source: [system] Destination table: [name] Natural keys: [columns] Change detection available: [updated_at / CDC / none] Expected volume: [rows/day] Warehouse: [Snowflake / BigQuery / Redshift / Databricks] Deletes in source: [hard deletes / soft deletes / unknown] Produce: 1. Recommended pattern (append-only, merge/upsert, snapshot, CDC) 2. Why it fits this source 3. SQL or dbt incremental model sketch 4. Watermark / high-water-mark handling, including ties and clock skew 5. Backfill plan and late data handling 6. Failure recovery steps 7. What breaks if deletes or late updates were misstated Do not use MAX(updated_at) as a watermark without calling out duplicates and failed batches.