SQL Patterns

SQL Patterns

Intro

SQL rewards thinking in sets and reading execution plans. Build queries incrementally from the core logic outward, lean on CTEs and window functions for clarity, and verify cost with EXPLAIN before optimizing anything.

Overview

Choose the right join

• INNER JOIN — only matching rows in both tables (default, most common).
• LEFT JOIN — all rows from the left table, NULLs where right has no match.
• RIGHT JOIN — rare; rewrite as LEFT JOIN for clarity.
• FULL OUTER JOIN — all rows from both, NULLs on the unmatched side.
• CROSS JOIN — cartesian product; use intentionally for generating combinations.

CTEs over nested subqueries

CTEs (WITH clauses) read top-to-bottom and let each step name its intent. Prefer them over nested subqueries when a query has more than two levels.

WITH active_users AS (
    SELECT user_id, MAX(login_at) AS last_login
    FROM logins
    WHERE login_at > CURRENT_DATE - INTERVAL '30 days'
    GROUP BY user_id
),
user_orders AS (
    SELECT user_id, COUNT(*) AS order_count, SUM(total) AS revenue
    FROM orders
    GROUP BY user_id
)
SELECT u.name, au.last_login, uo.order_count, uo.revenue
FROM users u
JOIN active_users au USING (user_id)
LEFT JOIN user_orders uo USING (user_id);

Window functions

Use windows for rankings, running totals, and per-row comparisons without collapsing rows:

SELECT
    department,
    employee,
    salary,
    RANK() OVER (PARTITION BY department ORDER BY salary DESC) AS dept_rank,
    salary - LAG(salary) OVER (PARTITION BY department ORDER BY salary) AS diff_from_prev,
    SUM(salary) OVER (PARTITION BY department) AS dept_total
FROM employees;

Remember the evaluation order: WHERE filters rows, GROUP BY groups them, HAVING filters groups, then window functions run after HAVING.

Optimization workflow

1. Read the EXPLAIN plan. Look for sequential scans on large tables, nested loops on big result sets, sort operations without indexes.
2. Check indexes. Ensure WHERE, JOIN, and ORDER BY columns are covered.
3. Reduce data early. Filter in WHERE before joining; join on indexed columns.
4. Rewrite anti-patterns. Functions on indexed columns, correlated subqueries, NOT IN with NULLs, implicit type casts — all defeat indexes.

Indexing strategy

• B-tree (default) — equality, range, ORDER BY, LIKE 'prefix%'.
• Hash — equality only, smaller than B-tree (Postgres).
• GIN — full-text search, JSONB containment, array operations.
• GiST — geometric data, range types, nearest-neighbor.
• Composite — column order matters (leftmost prefix rule).
• Partial — WHERE is_active = true for selective predicates.
• Covering — INCLUDE columns to enable index-only scans.

See references/schema-design.md for full schema-design patterns and references/query-patterns.md for query recipes.

Gotchas

Agent-specific failure modes — provider-neutral pause-and-self-check items:

• SELECT * in production queries. Selecting all columns returns data the application doesn't use, breaks when columns are added or removed, and prevents index-only scans. Always name the columns you need explicitly so the query is stable, the network payload is minimal, and the planner can use covering indexes.
• Functions on indexed columns in WHERE clauses. Writing WHERE LOWER(email) = ? or WHERE DATE(created_at) = ? forces the planner to evaluate the function on every row, making the index on email or created_at unusable. Use a functional index, normalize the data at write time, or rewrite the predicate to operate on the raw column.
• NOT IN with a subquery that can return NULLs. WHERE id NOT IN (SELECT user_id FROM orders) returns zero rows if any user_id in orders is NULL, because x NOT IN (..., NULL, ...) evaluates to NULL for every x. Use NOT EXISTS or filter out NULLs in the subquery explicitly.
• Correlated subqueries that execute once per row. A subquery in the SELECT list or WHERE clause that references the outer query's row runs once for every row in the outer result set — O(n) database round-trips. Rewrite as a JOIN or a lateral join so the planner can execute it once.
• Missing LIMIT on exploratory queries run against production. A full-table scan without a LIMIT clause can lock up a production database for minutes and saturate I/O. Add LIMIT 100 to any ad hoc query until you know the cardinality, and run exploratory queries against a read replica.
• Implicit type coercions in join conditions. Joining orders.user_id (INTEGER) to users.id (BIGINT) or VARCHAR forces the planner to cast one side for every comparison, often disabling index use. Match column types on both sides of every join; fix mismatches in the schema rather than relying on implicit casting.
• OFFSET pagination for deep pages. LIMIT 20 OFFSET 10000 forces the database to scan and discard 10,000 rows to return 20. For large offsets, performance degrades linearly. Use keyset (cursor) pagination — WHERE id > last_seen_id ORDER BY id LIMIT 20 — which uses the index and runs in constant time regardless of page depth.

Full reference

Pagination: keyset over offset

Offset pagination is simple but degrades on deep pages because the database must read and discard the skipped rows. Keyset (cursor) pagination is constant-time at any depth:

-- First page
SELECT id, name, created_at
FROM products
ORDER BY created_at DESC, id DESC
LIMIT 20;

-- Next page (using last row's values as cursor)
SELECT id, name, created_at
FROM products
WHERE (created_at, id) < ('2025-06-15 10:30:00', 4502)
ORDER BY created_at DESC, id DESC
LIMIT 20;

Add a composite index on (created_at DESC, id DESC) to support the sort.

Deduplication

Keep the latest row per key with a window function:

WITH ranked AS (
    SELECT *,
        ROW_NUMBER() OVER (
            PARTITION BY email ORDER BY updated_at DESC
        ) AS rn
    FROM contacts
)
SELECT * FROM ranked WHERE rn = 1;

Or delete duplicates in place:

DELETE FROM contacts c
USING (
    SELECT email, MAX(updated_at) AS keep_date
    FROM contacts GROUP BY email HAVING COUNT(*) > 1
) dupes
WHERE c.email = dupes.email AND c.updated_at < dupes.keep_date;

Pivot via conditional aggregation

SELECT
    department,
    COUNT(*) FILTER (WHERE status = 'active')   AS active_count,
    COUNT(*) FILTER (WHERE status = 'inactive') AS inactive_count,
    COUNT(*) FILTER (WHERE status = 'pending')  AS pending_count
FROM employees
GROUP BY department;

For engines without FILTER, fall back to SUM(CASE WHEN ... THEN 1 ELSE 0 END).

Recursive CTEs

Tree traversal with a depth limit to prevent runaway recursion on cyclic data:

WITH RECURSIVE org_tree AS (
    SELECT id, name, manager_id, 1 AS depth, ARRAY[name] AS path
    FROM employees
    WHERE manager_id IS NULL

    UNION ALL

    SELECT e.id, e.name, e.manager_id, t.depth + 1, t.path || e.name
    FROM employees e
    JOIN org_tree t ON e.manager_id = t.id
    WHERE t.depth < 10
)
SELECT id, name, depth, array_to_string(path, ' > ') AS chain
FROM org_tree
ORDER BY path;

Upsert

INSERT INTO metrics (sensor_id, reading_date, value)
VALUES ('temp-01', '2025-06-15', 23.5)
ON CONFLICT (sensor_id, reading_date)
DO UPDATE SET value = EXCLUDED.value, updated_at = NOW();

MySQL uses ON DUPLICATE KEY UPDATE; SQLite uses similar ON CONFLICT syntax.

Anti-join (rows in A not in B)

Prefer NOT EXISTS over NOT IN — NOT IN returns no rows if any NULL appears in the subquery result:

SELECT c.id, c.name
FROM customers c
WHERE NOT EXISTS (
    SELECT 1 FROM orders o WHERE o.customer_id = c.id
);

Composite index column order

The leftmost-prefix rule determines which queries an index supports. Given (status, created_at, category_id):

WHERE status = 'active'                                      -- yes
WHERE status = 'active' AND created_at > '2025-01-01'        -- yes
WHERE status = 'active' AND created_at > '2025-01-01'
  AND category_id = 5                                        -- yes
WHERE created_at > '2025-01-01'                              -- no
WHERE category_id = 5                                        -- no

Put equality columns first, then range columns, then sort columns.

Standard table boilerplate

CREATE TABLE example (
    id BIGINT GENERATED ALWAYS AS IDENTITY PRIMARY KEY,
    -- domain columns here
    created_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    updated_at TIMESTAMPTZ NOT NULL DEFAULT NOW()
);

Use TIMESTAMPTZ, never TIMESTAMP without zone. Name constraints explicitly (chk_orders_positive_total, uq_users_email) so error messages are readable.

Transactions and concurrency

• Use explicit transactions for multi-statement operations.
• Pick the right isolation level: READ COMMITTED (default, fine for most), REPEATABLE READ (consistent snapshots), SERIALIZABLE (full isolation, retry on conflict).
• Keep transactions short — never wait for user input inside one.
• Use SELECT ... FOR UPDATE when reading rows you intend to modify.
• Handle deadlocks with retry logic, not longer timeouts.

Anti-patterns

• SELECT * in production — fetches columns the caller does not need and breaks when schemas evolve.
• Functions on indexed columns — WHERE YEAR(created_at) = 2025 defeats the index; use a range instead.
• NOT IN with nullable subqueries — use NOT EXISTS.
• Correlated subqueries that run per row — rewrite as a join or CTE.
• Missing LIMIT on exploratory queries — easy way to OOM a client.
• Implicit type conversions — comparing a BIGINT column to a string literal can prevent index use.
• Storing money as floats — use DECIMAL(19,4) or integer cents.
• Missing foreign keys "for flexibility" — let the database enforce integrity.