Debug Slow Queries

Quick Start

Goal: go from "slow" to a verified fix with minimal risk.

When to use this skill

• p95 latency spikes, timeouts, or sudden cpu growth tied to database load
• a deploy introduces query regression
• tables grow and plans flip (index → seq scan, join strategy changes)

inputs to collect (before changing anything)

• database engine + version
• query text + typical parameters (not placeholders only)
• runtime evidence: p50/p95, rows returned, timeouts, error logs
• table sizes / row counts (approx)
• relevant indexes + constraints
• environment: prod vs staging, read replica vs primary

Workflow (default)

1. reproduce or capture
   • reproduce in staging with production-like data, or capture the exact query + params from logs
2. baseline
   • record current p50/p95 and the explain plan
3. read the plan
   • identify where time is spent (scan, join, sort, aggregate)
4. pick the cheapest safe fix
   • rewrite query → add/adjust index → update stats → change schema
5. verify
   • re-run explain, compare timing, validate correctness
6. roll out safely
   • ship behind a flag if needed, monitor, keep rollback path

Explain plan patterns (what to look for)

scans

• sequential scan on large table
  • likely missing selective predicate index, or predicate is not sargable
• index scan returning many rows
  • index exists but selectivity is poor; may need different leading columns or a partial index

joins

• nested loop with huge inner iterations
  • inner side needs an index on join key, or join order needs improvement
• hash join spilling
  • work_mem/memory pressure or too many rows; reduce join input with earlier filters

sorts / aggregates

• sort on large result
  • add index to match order by, or paginate differently
• group by on many rows
  • pre-filter, pre-aggregate, or add covering index for grouping columns

Engine-specific commands

PostgreSQL

get plan + timing:

EXPLAIN (ANALYZE, BUFFERS, VERBOSE) <query>;

check table stats freshness:

SELECT relname, n_live_tup, last_analyze, last_autoanalyze
FROM pg_stat_user_tables
ORDER BY n_live_tup DESC;

see indexes:

SELECT tablename, indexname, indexdef
FROM pg_indexes
WHERE schemaname = 'public' AND tablename = 'your_table';

safe index creation:

CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_name ON your_table (col1, col2);

MySQL

get plan:

EXPLAIN <query>;

get runtime plan details (mysql 8):

EXPLAIN ANALYZE <query>;

see indexes:

SHOW INDEX FROM your_table;

SQLite

get plan:

EXPLAIN QUERY PLAN <query>;

Fix toolbox (ordered by typical safety)

1) query rewrites (often best first)

• limit early
  • move filters into subqueries/ctes only if they reduce rows before joins
• avoid non-sargable predicates
  • avoid wrapping indexed columns in functions in where clauses
• replace select *
  • reduce io and sort payload
• avoid offset pagination at scale
  • use keyset pagination when possible

2) index changes

Use this checklist:

• does the where clause filter on a selective column?
• does the join predicate have an index on the inner side?
• does order by match an index prefix?
• is a composite index needed (leading columns matter)?
• can a partial index reduce size (postgres)?

3) statistics / maintenance

• analyze / vacuum (postgres) or optimize/analyze table (mysql) can fix bad estimates
• confirm cardinality estimates vs actual rows in explain analyze

4) schema-level changes (highest cost)

• denormalize only after proving query/index fixes are insufficient
• consider materialized views / summary tables for heavy aggregates

Output format (copy/paste)

## slow query report

### symptom

- endpoint/job:
- p50/p95:
- query:

### evidence

- explain plan notes:
- row counts:
- indexes involved:

### root cause

- primary bottleneck:
- why it happens:

### fix

- change:
- risk:
- rollback:

### verification

- before:
- after:
- correctness checks: