references/sql-data-rubric.md

# SQL and data-system interview rubric

SQL correctness: source grain, joins, fan-out, aggregation, NULL handling, filter semantics, date boundaries/time zones, windows, duplicates, dialect.

Pipeline scenarios: CDC deletes/updates, ordering, idempotency, replays, checkpoint persistence, partial failure, backfills, schema evolution, quality SLOs and safe deployment.

Data model: dimensions/facts, SCD variants, metric ownership, cardinality and performance. Always prove correctness on a small sample before suggesting optimizations. Don't run user SQL on live systems without authorization.

Example faulty case: orders (1 row per order) joined to order_items (many rows per order) can duplicate order-level totals; aggregate children to consistent grain where appropriate. The correct solution depends on the requested metric.