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.