SQL Architect

SQL Architecture & Database Standards

Targets PostgreSQL 18 as the primary engine; MySQL 9 and SQLite 3.53 noted where they differ. See STACK.md for pinned tool versions.

1. Schema design

• Primary key: id UUID PRIMARY KEY DEFAULT uuidv7() on every table. UUID v7 is sortable, distributed-friendly, and doesn't leak counts via URLs. PG 18 has native uuidv7(); on earlier engines use an extension or app-generated UUID v7.
• Natural identifiers stay UNIQUE: the surrogate id is for joins; domain meaning lives in UNIQUE constraints (email, slug, iso_code).
• Foreign keys: always declared at the DB level (REFERENCES other(id) ON DELETE RESTRICT by default). Never enforce relationships in app code alone.
• NOT NULL by default. Make NULL an explicit, justified choice — every nullable column should have a documented reason.
• CHECK constraints: push invariants into the DB (CHECK (price >= 0), CHECK (status IN (...))). They survive bad code paths.
• Audit columns (opt-in): created_at + updated_at (timestamptz NOT NULL DEFAULT now()) on tables whose rows change after creation. Skip on pure lookup tables (countries, currencies) and event/log tables (where event_at alone is enough).
• Soft delete (default): deleted_at timestamptz NULL. Filter via partial indexes (CREATE INDEX ... WHERE deleted_at IS NULL) so queries stay fast. Use hard delete only when GDPR/compliance requires it, or for append-only event tables where deletion never makes sense.

2. Naming conventions

• Tables: snake_case, plural (users, order_items).
• Columns: snake_case, singular (email, created_at).
• PK: always id.
• FKs: <reftable_singular>_id (user_id, order_id).
• Indexes (PostgreSQL): <table>_<cols>_idx; unique <table>_<cols>_key; partial <table>_<cols>_active_idx — matches Postgres's own auto-generated names. MySQL and SQLite differ — see ENGINES.md.
• Constraints: <table>_<purpose>_check, <table>_<purpose>_fkey.
• Sequences: PG generates them; rename only if necessary, never reference manually — UUID v7 PKs avoid the need.

3. Indexing strategy

• Index every FK column — JOINs without indexes are silent killers.
• B-tree is the default. GIN for arrays / JSONB containment. GiST for ranges / geometry. BRIN for huge append-only tables sorted by time.
• Composite index column order: most-selective column first, then by query pattern. The same composite index can serve WHERE a = ? and WHERE a = ? AND b = ? — but not WHERE b = ? alone.
• Partial indexes for soft-deleted rows (... WHERE deleted_at IS NULL) and other common filters.
• Don't over-index. Each index slows writes and consumes RAM. Add one only when EXPLAIN ANALYZE proves a real query needs it.
• Drop unused indexes. PG: query pg_stat_user_indexes periodically; remove any with idx_scan = 0.

4. Query patterns

• Parameter binding always. WHERE id = $1 — never string-concatenate user input.
• Explicit columns. SELECT id, email, created_at FROM users — never SELECT * in production code (breaks on schema additions, returns extra bytes).
• RETURNING on writes. INSERT ... RETURNING id, created_at saves a round-trip and surfaces server defaults.
• ON CONFLICT for upserts. INSERT ... ON CONFLICT (email) DO UPDATE SET ... — atomic, no race.
• Pagination = cursor, not OFFSET. WHERE (created_at, id) < ($1, $2) ORDER BY created_at DESC, id DESC LIMIT 20. OFFSET is O(N) on every page.
• CTEs (WITH) for readability, but be aware PG ≤ 11 materializes them (perf cliff). PG 12+ inlines unless MATERIALIZED is forced.

5. JOINs & N+1 prevention

• N+1 is the single most common perf bug. One query that fetches parents followed by N queries for children = always wrong.
• Detect it: log every SQL statement in dev with timings; any list endpoint emitting more than 2–3 queries per request is suspect.
• Fix it: single JOIN, or batched WHERE id IN (...) for the children, or DataLoader-style batchers if behind an API layer.
• JOIN vs subquery: different shapes, similar plans in modern PG — let EXPLAIN ANALYZE decide. Prefer JOINs when both sides return many rows; correlated subqueries when the inner table is small or you want a scalar.
• LEFT JOIN only when you genuinely need rows from the left without a match. A LEFT JOIN filtered by WHERE right.x = ? silently becomes an INNER JOIN — clearer to write it that way.

6. Transactions & isolation

• Wrap related writes in a transaction. Anything that must succeed or fail together.
• Isolation default: READ COMMITTED (PG default). Bump to REPEATABLE READ for multi-statement reads that need a consistent snapshot, and SERIALIZABLE when reads-then-writes need true linearizability — handle 40001 (serialization failure) by retrying.
• Keep transactions short. Holding a transaction across a network call or user input is a deadlock waiting to happen.
• Advisory locks (pg_advisory_lock) for distributed coordination that doesn't fit a row-level lock — singleton workers, leader election, idempotent jobs.
• Roll back explicitly on error. Application drivers should bracket every transaction in begin/commit/rollback — never leak open transactions.

7. Migrations

• Forward-only. No down migrations in production. Revert by writing a new forward migration. Down migrations rot fast and lie about what they restore.
• Idempotent where possible: CREATE TABLE IF NOT EXISTS, ADD COLUMN IF NOT EXISTS. Lets a half-applied migration re-run safely.
• One change per migration. Easier to review, easier to revert, atomic in CI.
• Naming: NNNNNN_snake_case_description.sql (e.g. 000023_add_orders_paid_at_index.sql). Number monotonically; pad enough to last.
• Large-table changes: add column as nullable first, backfill in batches (separate migration or async job), then add NOT NULL / index. Single ALTER TABLE ... NOT NULL on a billion-row table will lock writes for minutes.
• Tools: golang-migrate (Go) and alembic (Python) — both run raw .sql files. STACK.md pins versions.

8. Security

• Parameter binding always. Restated because it's the single most common SQL injection vector.
• Least-privilege roles. The app's runtime role can SELECT/INSERT/UPDATE/DELETE but should not CREATE/DROP/ALTER. Migrations run as a separate, elevated role.
• Row-Level Security (PG) for multi-tenant: ENABLE ROW LEVEL SECURITY + policies keyed on a session variable holding tenant_id. Defense in depth against app-layer mistakes.
• No secrets in connection strings stored in repo. Pull from env / secret manager. PG supports ~/.pgpass for local dev.
• Encrypted at rest is a deployment concern, but flag it: enable at the volume / disk layer.

9. Performance & EXPLAIN

• EXPLAIN ANALYZE is the only honest perf tool. Don't optimize without it.
• Read the plan top-to-bottom. Look for Seq Scan on large tables (missing index?), high actual rows vs plan rows (stale stats — run ANALYZE), nested loops over huge inputs (wrong index).
• Slow query logging: log_min_duration_statement = 200 (ms) in dev. Tighter in prod, but at least log the 99th percentile offenders.
• Connection pooling. Postgres pays a real cost per connection. Use pgbouncer (transaction-mode for OLTP) in front; in-app pools (e.g. pgx, psycopg) for finer control.
• VACUUM ANALYZE runs automatically (autovacuum) — be aware of bloat on heavy UPDATE/DELETE tables and tune autovacuum thresholds for them.

10. JSON columns — when (and when not)

• Use JSONB only for genuinely schemaless payloads: raw webhook bodies, opaque third-party event blobs, user-defined opaque config.
• Never for structured domain data. If you find yourself querying data->'order'->>'status' repeatedly, it's a column, not a JSON field. Normalize.
• Index with GIN if you must query JSONB: CREATE INDEX ... USING GIN (data jsonb_path_ops) for containment queries.

11. Engine-specific guidance

PostgreSQL 18 is the primary target. Per-engine references are split out to keep this skill focused:

• PostgreSQL-specific types and features (timestamptz, text, citext, generated columns, partitioning, RLS, native uuidv7(), etc.) — see POSTGRES.md.
• MySQL 9 and SQLite 3.53 notes (charset, primary key strategies, JSON support, concurrency caveats) — see ENGINES.md.