Postgresdb Agent Skill

PostgreSQL — schema, indexing, queries, ops

Engine-level PostgreSQL 16 guidance: design correct schemas, pick the right index, read EXPLAIN and

fix slow SQL, run zero-downtime migrations, and operate/secure the database. Tooling-agnostic; every

example is runnable.

New, non-trivial feature with no approved spec + plan under 02-DOCS/wiki/sdd/? Hand off to

specify before writing feature code (method: sdd); build

straight from here only for a genuinely one-line, low-risk change.

Deep dives: schema-and-indexing (types, constraints, every index

kind, bloat) · query-optimization (EXPLAIN, joins, concurrency,

JSONB/FTS/pgvector) · migrations (zero-downtime DDL, per-ORM) ·

operations-and-security (roles, RLS, pooling, vacuum,

partitioning, backups).

Not this skill. ORM-API ergonomics (Prisma updateMany count trap, SQLAlchemy session lifecycle)

and per-runner migration wiring → that tool's own docs; this skill owns the **SQL the ORM emits and

the engine behavior underneath**. Other engines → mysql,

sqlite-turso, clickhouse-analytics

(different MVCC, locking, planner). App-layer caching / Redis / Kafka as products are out — only

Postgres-as-queue via SKIP LOCKED is in scope. Cloud-vendor console clicks →

deployment; we give the SQL and params, not the RDS/Cloud SQL UI path.

Decision rules

Fast lookups; runnable DDL lives in the references.

Pick the column type

| Use case | Correct type | Avoid | Why |

| --- | --- | --- | --- |

| Surrogate PK (internal) | bigint GENERATED ALWAYS AS IDENTITY | serial, int | identity is SQL-standard, no sequence-ownership gotchas; bigint avoids 2.1B overflow |

| Surrogate PK (public/distributed) | uuid v7 | uuid v4 | v7 is time-ordered → less B-tree fragmentation than random v4 |

| Natural text id (slug, sku) | text + UNIQUE + CHECK | varchar(n) | length via CHECK; no rewrite to widen later |

| Money / exact decimal | numeric(19,4) | float8, money | binary floats drift; money has locale issues |

| Timestamp (event) | timestamptz | timestamp | stores a UTC instant; naive timestamp loses zone |

| Duration | interval | int seconds | self-documenting, arithmetic-safe |

| Small closed set, stable | enum | text w/o CHECK | type safety; but see lookup-table note |

| Evolving set, joinable | lookup table + FK | enum | ALTER TYPE ... ADD VALUE is awkward; FK gives joins + soft-retire |

| Flag | boolean | int, varchar | three-valued NULL still possible — add NOT NULL DEFAULT |

| Tags (read-mostly) | text[] + GIN | comma string | array ops + GIN containment |

| Tags (relational) | join table | text[] | when you need FK integrity / per-tag rows |

| Semi-structured | jsonb | json, text | binary, indexable, dedup keys; promote hot keys to columns |

| IP / CIDR | inet / cidr | text | validation + operators |

| Time range (booking) | tstzrange + GiST | two columns | && overlap + exclusion constraint |

Pick the index

| Access pattern | Index | DDL | Notes |

| --- | --- | --- | --- |

| = / < > / range / ORDER BY | btree (default) | CREATE INDEX ix_orders_status ON orders (status) | also enforces uniqueness |

| LIKE 'prefix%' | btree + text_pattern_ops | CREATE INDEX ix_users_email_pat ON users (email text_pattern_ops) | only for C-locale/prefix; not %suffix |

| Case-insensitive eq | expr index or citext | CREATE INDEX ix_users_lemail ON users (lower(email)) | query must use lower(email) too |

| @> jsonb / array containment | GIN | CREATE INDEX ix_orders_meta ON orders USING gin (meta) | jsonb_path_ops if only @> |

| Full-text @@ | GIN on tsvector | CREATE INDEX ix_orders_search ON orders USING gin (search) | index a generated tsvector column |

| Range overlap / exclusion / geo | GiST | CREATE INDEX ix_book_during ON bookings USING gist (during) | also PostGIS geometry |

| Huge append-only time-series | BRIN | CREATE INDEX ix_events_ts ON events USING brin (created_at) | needs physical correlation |

| Vector similarity | hnsw (pgvector) | CREATE INDEX ix_docs_embed ON docs USING hnsw (embedding vector_cosine_ops) | see query-optimization |

| Dedup only | unique btree | CREATE UNIQUE INDEX uq_users_email ON users (email) | constraint = index |

Hash indexes: almost never — equality-only, no multicolumn, rarely beats btree even though WAL-logged

since PG10.

Read the plan before adding an index — EXPLAIN (ANALYZE, BUFFERS) or it didn't happen. An index

the planner never picks is pure write tax on every insert and update, forever.

When NOT to add an index

• Low-selectivity boolean / status with few distinct values (planner ignores it; seq scan wins).
• Tiny tables (a seq scan reads one page; the index adds maintenance for nothing).
• Write-heavy columns rarely filtered — every index is a write tax.
• A column already the left prefix of an existing composite index.
• Redundant with a UNIQUE constraint — the constraint already created an index.

Copy-paste patterns

Canonical table (types + constraints + identity)

-- GOOD: identity PK, public uuid, FK with action, numeric money, timestamptz,
-- status via lookup FK, generated tsvector, CHECK constraints.
CREATE TABLE orders (
    id           bigint GENERATED ALWAYS AS IDENTITY PRIMARY KEY,
    public_id    uuid NOT NULL DEFAULT gen_random_uuid(),  -- v4; see schema ref for v7
    user_id      bigint NOT NULL REFERENCES users (id) ON DELETE RESTRICT,
    status       text NOT NULL REFERENCES order_statuses (code) ON UPDATE CASCADE,
    amount       numeric(19,4) NOT NULL CHECK (amount >= 0),
    currency     text NOT NULL CHECK (length(currency) = 3),
    note         text,
    search       tsvector GENERATED ALWAYS AS (to_tsvector('simple', coalesce(note, ''))) STORED,
    created_at   timestamptz NOT NULL DEFAULT now(),
    updated_at   timestamptz NOT NULL DEFAULT now(),
    UNIQUE (public_id)
);
CREATE INDEX ix_orders_user_id ON orders (user_id);     -- FKs are NOT auto-indexed
CREATE INDEX ix_orders_status  ON orders (status);

-- BAD: every column is a future migration or a bug.
CREATE TABLE orders (
    id          serial PRIMARY KEY,            -- sequence-ownership gotchas; use IDENTITY
    user_id     int REFERENCES users(id),      -- int overflows at 2.1B; no FK index
    status      varchar(20),                   -- length hack; no constraint on values
    amount      float,                         -- money drift
    created_at  timestamp DEFAULT now()        -- naive: loses the zone
);

updated_at is not auto-maintained — add a BEFORE UPDATE trigger (see schema ref) or set it in the

app; Postgres has no ON UPDATE clause.

Composite index column order

Equality columns first, then the range/sort column.

-- Query: WHERE user_id = $1 AND created_at >= $2 ORDER BY created_at DESC
-- GOOD: equality (user_id) then range (created_at)
CREATE INDEX ix_orders_user_created ON orders (user_id, created_at DESC);

-- BAD: range-first index cannot satisfy the equality efficiently for this query
CREATE INDEX ix_orders_created_user ON orders (created_at, user_id);

Confirm with EXPLAIN that the plan shows Index Cond: (user_id = ... AND created_at >= ...), not a

Filter:.

Partial + covering (INCLUDE) index

-- Partial: index only the rows you query (smaller, hotter)
CREATE INDEX ix_orders_active ON orders (user_id, created_at DESC)
WHERE status <> 'cancelled';

-- Covering: INCLUDE non-key columns to enable an index-only scan
CREATE INDEX ix_orders_user_cover ON orders (user_id) INCLUDE (amount, created_at);

Index-only scan requires a recently-VACUUMed table; confirm Heap Fetches: 0 in EXPLAIN (ANALYZE).

Keyset (cursor) pagination — not OFFSET

-- GOOD: keyset on a stable composite sort; uses ix_orders_user_created
SELECT id, amount, created_at
FROM orders
WHERE user_id = $1
  AND (created_at, id) < ($2, $3)   -- row-value comparator = last row of prev page
ORDER BY created_at DESC, id DESC
LIMIT 20;

-- BAD: OFFSET scans and discards 100000 rows every page (O(n))
SELECT * FROM orders WHERE user_id = $1 ORDER BY created_at DESC LIMIT 20 OFFSET 100000;

The index must match the ORDER BY direction exactly; include the tiebreaker (id).

UPSERT done right

-- GOOD: insert-or-update; EXCLUDED is the row that failed to insert
INSERT INTO inventory (sku, qty)
VALUES ($1, $2)
ON CONFLICT (sku)
DO UPDATE SET qty = inventory.qty + EXCLUDED.qty
WHERE inventory.qty + EXCLUDED.qty >= 0   -- guard
RETURNING id, qty;

-- DO NOTHING returns no row on conflict; wrap to always get the row:
WITH ins AS (
    INSERT INTO tags (name) VALUES ($1)
    ON CONFLICT (name) DO NOTHING
    RETURNING id
)
SELECT id FROM ins
UNION ALL
SELECT id FROM tags WHERE name = $1 LIMIT 1;

Queue with SKIP LOCKED

-- GOOD: contention-free job claim; concurrent workers never block each other
UPDATE jobs
SET status = 'processing', locked_at = now()
WHERE id = (
    SELECT id FROM jobs
    WHERE status = 'pending'
    ORDER BY created_at
    FOR UPDATE SKIP LOCKED
    LIMIT 1
)
RETURNING id, payload;

SKIP LOCKED skips rows another txn holds; FOR UPDATE alone would serialize all workers.

Kill the N+1

-- BAD: application loops, one query per order (N+1)
--   for o in orders: SELECT * FROM order_items WHERE order_id = o.id
-- GOOD: one set-based query
SELECT o.id, json_agg(i.*) AS items
FROM orders o
JOIN order_items i ON i.order_id = o.id
WHERE o.user_id = $1
GROUP BY o.id;

-- Top-N-per-group: JOIN LATERAL, not a window-filter scan
SELECT u.id, recent.*
FROM users u
JOIN LATERAL (
    SELECT id, amount, created_at FROM orders
    WHERE user_id = u.id ORDER BY created_at DESC LIMIT 3
) recent ON true;

An ORM emitting N queries is the same bug — fix it at the SQL boundary, not with a cache.

EXPLAIN, the right way

EXPLAIN (ANALYZE, BUFFERS, SETTINGS, FORMAT TEXT)
SELECT * FROM orders WHERE user_id = $1 AND created_at >= $2;

Read these four first:

• Estimated vs actual rows — a large gap means stale stats; run ANALYZE.
• The most expensive node — highest actual time × loops.
• Seq Scan on a big table where you expected an index.
• Rows Removed by Filter — the predicate was not pushed to an index.

Full method in query-optimization.

DDL on a live table

Full sequences (zero-downtime expand-contract, batched backfills, per-ORM runners) in

migrations. These four are absolute because each one is a lock you cannot

take back once traffic is on the table:

• Index creation on a live table is always CONCURRENTLY — plain CREATE INDEX holds ACCESS

EXCLUSIVE for the entire build and blocks every writer. It therefore cannot run inside a txn.

• Never ADD COLUMN ... NOT NULL without a default/backfill plan, and never add a volatile

default (now(), gen_random_uuid()) on a large table without a batched backfill — a volatile

default rewrites the whole table under ACCESS EXCLUSIVE. A non-volatile constant is instant (PG11+).

• SET lock_timeout + SET statement_timeout around DDL on hot tables, so a blocked statement fails

fast instead of parking an ACCESS EXCLUSIVE request that every reader behind it then queues on.

• Migrations are forward-only in prod; never edit an applied migration — it has already run

somewhere, so the next environment replays a history that no longer matches the one in production.

Lock modes: CREATE INDEX CONCURRENTLY takes SHARE UPDATE EXCLUSIVE (allows writes); plain

CREATE INDEX, ALTER TABLE ... TYPE, ADD COLUMN with a volatile default, and VACUUM FULL take

ACCESS EXCLUSIVE (blocks everything). Full table in

migrations.

Anti-patterns

| Claim | Reality |

| --- | --- |

| "I'll add the FK index later, the query works now" | Unindexed FK = seq scan + heavy lock cascade on parent DELETE/UPDATE. Index it now. |

| "UUID PK is fine everywhere" | Random v4 fragments the B-tree and bloats WAL. Use IDENTITY internally or uuid v7. |

| "SELECT count(*) to check existence" | Counts the whole match. Use EXISTS (SELECT 1 ...). |

| "CTEs are just for readability" | Pre-12 they were optimization fences; PG12+ inlines unless MATERIALIZED. Know which you want. |

| "NOT IN (subquery)" | NULL-unsafe (one NULL → empty result) and slow. Use NOT EXISTS. |

| "Store money as float, round on display" | Silent drift across arithmetic. numeric(19,4). |

| "ADD COLUMN ... NOT NULL DEFAULT now()" | Volatile default rewrites the table under ACCESS EXCLUSIVE. Non-volatile constant is instant (PG11+). |

| "One big jsonb blob beats columns" | No constraints, no per-key stats, GIN bloat. Promote hot keys to typed columns. |

| "RLS is on, so the table is protected" | RLS is opt-in per table and the table owner bypasses it. Verify with a non-owner role; add FORCE ROW LEVEL SECURITY to cover the owner too. |

| "RLS policy calling auth.uid() per row" | Re-evaluated per row. Wrap: (SELECT auth.uid()) so it runs once. |

| "CREATE INDEX in the migration is fine" | Blocks writes for the whole build. CREATE INDEX CONCURRENTLY (outside a txn). |

| "VACUUM FULL will fix bloat" | Takes ACCESS EXCLUSIVE, rewrites the table. Use autovacuum tuning / REINDEX CONCURRENTLY. |

Quick reference

Isolation levels

| Level | Prevents | Use when | Note |

| --- | --- | --- | --- |

| Read Committed (default) | dirty reads | most OLTP | each statement sees a fresh snapshot |

| Repeatable Read | + non-repeatable / phantom (snapshot) | multi-statement consistent read | may raise 40001; retry |

| Serializable (SSI) | + write skew | invariants across rows | retry 40001 with backoff |

Retry the txn on SQLSTATE 40001 (serialization_failure) and 40P01 (deadlock_detected).

Diagnostic one-liners

-- Unindexed foreign keys
SELECT c.conrelid::regclass AS tbl, a.attname AS col
FROM pg_constraint c
JOIN pg_attribute a ON a.attrelid = c.conrelid AND a.attnum = ANY (c.conkey)
WHERE c.contype = 'f'
  AND NOT EXISTS (
    SELECT 1 FROM pg_index i
    WHERE i.indrelid = c.conrelid AND (i.indkey::int2[])[0] = a.attnum
  );

-- Top slow queries (needs pg_stat_statements)
SELECT calls, round(mean_exec_time::numeric, 2) AS mean_ms,
       round(total_exec_time::numeric, 2) AS total_ms, query
FROM pg_stat_statements ORDER BY total_exec_time DESC LIMIT 20;

-- Dead tuples / bloat candidates
SELECT relname, n_dead_tup, n_live_tup, last_autovacuum
FROM pg_stat_user_tables WHERE n_dead_tup > 1000 ORDER BY n_dead_tup DESC;

-- Blocking locks
SELECT blocked.pid AS blocked_pid, blocking.pid AS blocking_pid, blocked.query AS blocked_query
FROM pg_stat_activity blocked
JOIN pg_locks bl ON bl.pid = blocked.pid AND NOT bl.granted
JOIN pg_locks gl ON gl.locktype = bl.locktype AND gl.database IS NOT DISTINCT FROM bl.database
  AND gl.relation IS NOT DISTINCT FROM bl.relation AND gl.granted
JOIN pg_stat_activity blocking ON blocking.pid = gl.pid;

-- Cache hit ratio (aim > 0.99)
SELECT sum(heap_blks_hit) / nullif(sum(heap_blks_hit + heap_blks_read), 0) AS ratio
FROM pg_statio_user_tables;

-- Unused indexes
SELECT relname, indexrelname, idx_scan
FROM pg_stat_user_indexes WHERE idx_scan = 0 ORDER BY relname;

Verify

Run scripts/verify.sh from your project root: it lints discovered SQL with sqlfluff (if

configured), syntax-sanity-checks migration files (the quote/paren balance check is dollar-quote and

block-comment aware), flags foot-guns (CREATE INDEX without CONCURRENTLY in a migration,

ADD COLUMN ... NOT NULL without DEFAULT, VACUUM FULL), and — only if DATABASE_URL and psql

are present — checks that pg_stat_statements is enabled. It exits non-zero only on a real

sqlfluff lint error; everything else (missing tools, heuristic warnings, DB unreachable) is advisory

[skip]/[warn]. Runs on stock macOS bash 3.2; never writes, never connects without DATABASE_URL.

Project grounding

In a project with a 02-DOCS/ layer (the harness Karpathy wiki), read

02-DOCS/wiki/stack/postgresdb.md first and stay consistent with it. Missing or stale? Write the

project's real choices there — schema and naming conventions, migration tool, indexing/partitioning

decisions, pooling setup, RLS policies — index it in 02-DOCS/wiki/index.md (the Knowledge map; root

CLAUDE.md keeps only a pointer), and bump its Updated date in the same change, so the next agent

inherits the conventions instead of re-deriving them. No 02-DOCS/ layer? Skip silently (optionally

suggest harness) — technical conventions are *recorded, not gated*; never block the task on this.