SQL Query Optimizer

SQL Query Optimizer

Make queries fast by understanding the plan, not by guessing. The costly mistake this skill prevents is the shotgun index: adding indexes on instinct, which slows every write and bloats storage while the actual cost driver - a non-sargable predicate, a mis-ordered composite, an N+1 in the ORM - survives untouched.

Operating procedure

Diagnosis strictly precedes any fix, because the right fix depends on which cost driver the plan reveals.

Step 1: gather inputs

Collect (label guesses as guesses):

1. The query text and, if ORM-generated, the ORM code that emits it.
2. EXPLAIN ANALYZE output (or EXPLAIN if production-unsafe to execute; note that estimates only).
3. Approximate row counts of the tables involved, and existing indexes (\d table or equivalent).
4. Current latency, target latency, and how often the query runs. A 200 ms query at 500 QPS matters more than a 5 s nightly report.
5. Write volume on the tables - this prices any new index.

Step 2: read the plan for cost drivers

Scan for these signals, in rough order of frequency:

• Sequential scan on a large table (over ~100k rows) under a selective WHERE - the classic missing-index signature. A seq scan on a small table is fine and often optimal; do not "fix" it.
• Estimated vs actual rows off by 10x or more - stale statistics or correlated columns; run ANALYZE (or create extended statistics) before trusting any other conclusion from the plan.
• Nested loop whose inner side executes thousands of times (loops=N large) over a non-trivial row count - join order or missing join index.
• Sort spilling to disk (Sort Method: external merge) - the sort exceeds work_mem; fix with an index providing the order, or raise work_mem only if the sort is genuinely necessary.
• Repeated subquery execution - a correlated subquery in the plan as a per-row SubPlan.
• Filter applied after an index scan removes most rows (Rows Removed by Filter large) - the index does not cover the real predicate.

Step 3: check for N+1 before optimizing the single query

Symptoms: one query is fast in isolation but the endpoint is slow; the DB log shows the same query shape hundreds of times per request with only the ID changing; latency scales with result count. If present, the fix is in the application (batch/JOIN/preload), not in this query - route to n-plus-one-hunter and stop tuning a query that should not run N times.

Step 4: apply the matching fix

• Missing index: B-tree on the columns in WHERE/JOIN/ORDER BY. Composite indexes follow the left-to-right rule - a query must use a leftmost prefix; order columns by the access pattern (equality columns first, then the range or sort column) and selectivity. For a whole-schema indexing strategy, route to index-advisor.
• Non-sargable predicate: never wrap the indexed column in a function. Rewrite WHERE date(created_at) = '...' as WHERE created_at >= '...' AND created_at < '...' + 1 day. Same for LOWER(email) = ... (use a functional index or citext) and leading-wildcard LIKE.
• SELECT *: project only needed columns so a covering index can satisfy the query without heap fetches.
• Correlated subquery: rewrite as a JOIN or a single grouped aggregate. Deep shape rewrites route to query-rewriter.
• OR across different columns: rewrite as a UNION of two independently-indexable queries.
• LIMIT with large OFFSET: replace with keyset pagination (WHERE id > last_seen ORDER BY id LIMIT n) - OFFSET still reads and discards every skipped row.

Step 5: verify and price it

Re-run EXPLAIN ANALYZE after the change. Claim a speedup only with before/after plan evidence. State the write-side cost of any new index: each index adds work to every INSERT/UPDATE on the table - on a write-heavy table (thousands of writes/minute), an index must earn its keep, and 5+ indexes on such a table deserves scrutiny.

Worked example

Before - 4.2 s on a 6M-row orders table:

SELECT * FROM orders
WHERE date(created_at) = '2026-06-01' AND status = 'shipped'
ORDER BY total DESC LIMIT 20;
-- Plan: Seq Scan on orders (rows=6,012,340), Filter removes 99.9%,
-- Sort Method: external merge  Disk: 214MB

Reasoning: date(created_at) defeats any index on created_at (non-sargable), so the planner seq-scans all 6M rows, then sorts the survivors on disk. Fix the predicate to a range, then give the planner one composite index serving the equality filter, the range, and the sort.

After - 8 ms:

CREATE INDEX idx_orders_status_created_total
  ON orders (status, created_at, total DESC);

SELECT id, customer_id, total, created_at FROM orders
WHERE status = 'shipped'
  AND created_at >= '2026-06-01' AND created_at < '2026-06-02'
ORDER BY total DESC LIMIT 20;
-- Plan: Index Scan using idx_orders_status_created_total (rows=20)

Equality column (status) leads, range (created_at) next, sort key last; the projection drops SELECT *. Write cost: one more index maintained per order write - acceptable unless orders sustains very high write rates.

Deliverable

Produce: the rewritten query, the index DDL, a one-or-two-sentence expected plan change, the before/after EXPLAIN evidence (or a note that verification is pending), and the write-side cost statement for each new index.

Do NOT

• Do not add an index before reading the plan - you may index around the real problem and tax every write for nothing.
• Do not trust a plan whose row estimates are 10x off from actuals; fix statistics first.
• Do not tune a query that runs N times per request when it should run once - that is an application bug (n-plus-one-hunter).
• Do not "optimize" seq scans on small tables or queries that return most of a table; a seq scan is the right plan there.
• Do not claim a speedup without post-change EXPLAIN ANALYZE evidence.

Quality bar

• The named cost driver is visible in the quoted plan, not asserted.
• The rewrite AND the index DDL are both shown, with composite column order justified.
• Expected plan change is stated in one or two sentences.
• Write-side cost of every new index is noted.
• Verification with EXPLAIN is done or explicitly flagged as the next step.