SQL Optimizer

SQL Optimizer Skill

A slow query almost always has a specific, findable cause — a missing index, a non-sargable predicate, a join that explodes rows, a scan that should be a seek. This skill diagnoses it: read what the query (and plan, if given) is actually doing, name the bottleneck, and produce a concrete rewrite plus the index / partition / structural changes — with the expected impact, not vague "add indexes" advice.

Required Inputs

Ask for these only if they aren't already provided:

• The query (and the engine — Postgres, BigQuery, Snowflake, MySQL… optimizations differ).
• The symptom — slow, expensive (bytes scanned), timing out, or just under review.
• Context if available — EXPLAIN/query plan, table sizes/row counts, existing indexes, partitioning/clustering.

Output Format

SQL Optimization: [query purpose]

1. What it's doing now — read the query (and plan): the scans, joins, sorts, and where the time/cost goes. Name the primary bottleneck (don't list ten micro-tweaks — find the one that matters).

2. The problems — ranked, each with why it's slow:

• Non-sargable predicates (functions on indexed columns, leading wildcards) → can't use an index.
• Missing/`wrong index or partition pruning; full scans where a seek is possible.
• Join issues — fan-out, wrong join order, missing join keys, SELECT * pulling everything.
• Sorts/spills, DISTINCT/GROUP BY on high-cardinality, correlated subqueries that should be joins.
• Engine-specific: BigQuery/Snowflake → bytes scanned (partition/cluster pruning), not row counts.

3. The fix — the rewritten query, plus the index / partition / clustering / materialization changes. Be specific (CREATE INDEX … ON … (cols), partition on event_date).

4. Expected impact — roughly what each change buys (seek vs. scan, pruning N% of partitions, removing a sort) and how to verify (re-run EXPLAIN, compare bytes/rows).

Quality Checks

• Names the single primary bottleneck, not a scattershot list
• Predicates are checked for sargability (no functions on indexed columns, no leading %)
• Index/partition recommendations are specific (exact columns), not "add an index"
• For columnar/cloud engines, addresses bytes scanned & pruning, not just row counts
• Provides the rewritten query and a way to verify the improvement

Anti-Patterns

• Do not say "add indexes" generically — name the columns and explain which predicate/join they serve
• Do not ignore the engine — Postgres index tuning and BigQuery partition pruning are different games
• Do not optimize a query that should be a model — repeated heavy logic belongs in a materialized/dbt model
• Do not wrap indexed columns in functions in the WHERE clause — it kills index usage (non-sargable)
• Do not recommend changes without an expected impact or a way to measure it

Based On

Query-optimization practice — sargability, index/partition pruning, join-order and fan-out, plan reading, columnar bytes-scanned tuning.