SKILL.md
Databricks SQL performance review
Use this skill when optimizing or reviewing SQL in Databricks SQL.
What to ask for (only if missing)
Ask up to 3 questions total:
- The query text (if not provided)
- The table(s) involved + their sizes (rough order of magnitude) OR the query profile / execution plan
- The desired result constraints (correctness, exactness, latency SLA)
If the user can’t provide sizes/plan, proceed with best-effort heuristics and call out assumptions.
Output format
Use the structure in assets/sql-review-output.md.
Checklist
Use references/sql-checklist.md to ensure you cover the common performance levers:
- predicate pushdown / partition pruning
- join strategy and join keys
- avoid
SELECT * - minimize shuffles / wide aggregations
- use correct data types and avoid implicit casts
- reduce data scanned (pre-filter, semi-joins, EXISTS)
Examples
User: “This query is slow in Databricks SQL. Can you optimize it?” (pastes query) Assistant: Provide issues, suggestions, and a rewritten query, plus next steps (EXPLAIN, add ZORDER, etc.).
Edge cases
- If the query is logically wrong (duplicates from joins, missing filters), fix correctness first.
- If tables are Delta: suggest partitioning/ZORDER/OPTIMIZE only if it matches query patterns.
- If the user is in a governed environment: avoid suggestions that require elevated permissions unless noted.