api-database-postgresql

PostgreSQL Patterns (node-postgres)

Quick Guide: Use the pg package (v8.x) for direct PostgreSQL access. Always use Pool -- never create individual Client instances in application code. Use parameterized queries ($1, $2) for ALL user input -- never interpolate strings into SQL. For transactions, check out a dedicated client with pool.connect() and use BEGIN/COMMIT/ROLLBACK in a try/catch/finally that always calls client.release(). Handle the pool error event to prevent process crashes from idle client errors. Use pg-query-stream for large result sets to avoid loading everything into memory.

---

<critical_requirements>

CRITICAL: Before Using This Skill

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST use parameterized queries ($1, $2, ...) for ALL values -- NEVER concatenate or interpolate user input into SQL strings)

(You MUST use Pool for all database access -- NEVER create standalone Client instances in application code)

(You MUST release clients back to the pool in a finally block after pool.connect() -- leaked clients exhaust the pool and hang the application)

(You MUST handle the error event on Pool instances -- unhandled idle client errors crash the Node.js process)

</critical_requirements>

---

Examples

• Core Patterns -- Pool setup, parameterized queries, type-safe results, error handling
• Transactions -- BEGIN/COMMIT/ROLLBACK, savepoints, retry logic, advisory locks
• Streaming -- Cursors, pg-query-stream, LISTEN/NOTIFY for real-time
• Advanced -- SSL/TLS, prepared statements, migrations, testing patterns

Additional resources:

• reference.md -- Pool options, error codes, QueryResult properties, production checklist

---

Auto-detection: PostgreSQL, pg, node-postgres, Pool, Client, pool.query, pool.connect, client.query, $1, parameterized query, BEGIN, COMMIT, ROLLBACK, LISTEN, NOTIFY, pg_notify, pg-query-stream, pg-cursor, Cursor, QueryResult, QueryResultRow, connectionString, PGHOST, PGDATABASE, unique_violation, 23505, deadlock, 40P01, advisory lock

When to use:

• Direct SQL queries against PostgreSQL (not behind an ORM)
• Connection pool management for Node.js/PostgreSQL applications
• Transactions spanning multiple queries that must be atomic
• Streaming large result sets without loading everything into memory
• Real-time change notifications via LISTEN/NOTIFY
• Integration testing with transaction rollback isolation

Key patterns covered:

• Pool configuration and lifecycle (creation, error handling, graceful shutdown)
• Parameterized queries with $1-style placeholders (SQL injection prevention)
• Type-safe query results using TypeScript generics
• Transaction management with dedicated clients
• Streaming with pg-cursor and pg-query-stream
• LISTEN/NOTIFY for real-time PostgreSQL event handling
• PostgreSQL error code handling (constraint violations, deadlocks, serialization failures)
• SSL/TLS connection configuration
• Testing with transaction rollback isolation

When NOT to use:

• You need an ORM or query builder -- use your ORM/query builder skill instead
• You need in-memory caching -- use a caching solution
• You need document storage without relational constraints -- use a document database
• Simple key-value lookups at sub-millisecond latency -- use an in-memory data store

---

Philosophy

pg (node-postgres) is a low-level PostgreSQL client that gives you full control over SQL, connections, and transactions. The core principle: write SQL directly, let PostgreSQL do the heavy lifting.

Core principles:

1. Pool, never Client -- Application code should always use Pool. The pool manages connections, handles reconnection, and prevents connection exhaustion. Use pool.query() for single queries, pool.connect() when you need a dedicated client (transactions).
2. Parameterized everything -- Never build SQL by string concatenation. Use $1, $2 placeholders. This prevents SQL injection AND enables PostgreSQL query plan caching.
3. Release in finally -- Any client obtained via pool.connect() must be released in a finally block. A leaked client sits checked out forever, and once max clients leak, the pool deadlocks.
4. Fail loudly -- Handle the pool's error event. Handle query errors with specific PostgreSQL error codes. Never swallow errors silently.
5. Stream large results -- Don't SELECT * a million rows into memory. Use pg-cursor or pg-query-stream for large result sets.

---

Core Patterns

Pattern 1: Pool Setup

Create a single pool per database at application startup. See examples/core.md for full configuration examples.

// ✅ Good Example - Pool with error handling
import pg from "pg";

const POOL_MAX_CLIENTS = 20;
const IDLE_TIMEOUT_MS = 30_000;
const CONNECTION_TIMEOUT_MS = 5_000;

function createPool(): pg.Pool {
  const pool = new pg.Pool({
    connectionString: process.env.DATABASE_URL,
    max: POOL_MAX_CLIENTS,
    idleTimeoutMillis: IDLE_TIMEOUT_MS,
    connectionTimeoutMillis: CONNECTION_TIMEOUT_MS,
  });

  pool.on("error", (err) => {
    console.error("Unexpected idle client error:", err.message);
  });

  return pool;
}

export { createPool };

Why good: Named constants for pool config, environment variable for connection string, error handler prevents process crash from idle client errors

// ❌ Bad Example - No pool, standalone client
import pg from "pg";

const client = new pg.Client("postgres://localhost/mydb");
await client.connect();
// One connection for entire app -- no pooling, no reconnection,
// no concurrency. If client disconnects, app crashes.

Why bad: Standalone Client has no connection pooling, no automatic reconnection, no concurrency -- every query blocks on a single connection

---

Pattern 2: Parameterized Queries

Always use $1-style placeholders. See examples/core.md for typed query helpers.

// ✅ Good Example - Parameterized query with typed result
interface UserRow {
  id: number;
  name: string;
  email: string;
}

const result = await pool.query<UserRow>(
  "SELECT id, name, email FROM users WHERE id = $1",
  [userId],
);

const user = result.rows[0]; // UserRow | undefined

Why good: $1 placeholder prevents SQL injection, generic <UserRow> types the rows array, result is properly typed

// ❌ Bad Example - String interpolation (SQL INJECTION!)
const result = await pool.query(
  `SELECT * FROM users WHERE name = '${userName}'`,
);
// userName = "'; DROP TABLE users; --" -> catastrophic

Why bad: String interpolation allows SQL injection, no type safety on result rows, SELECT * returns untyped columns

---

Pattern 3: Transactions

Use pool.connect() to get a dedicated client for the transaction. See examples/transactions.md for savepoints, retries, and advisory locks.

// ✅ Good Example - Transaction with proper cleanup
async function transferFunds(
  pool: pg.Pool,
  fromId: number,
  toId: number,
  amount: number,
): Promise<void> {
  const client = await pool.connect();
  try {
    await client.query("BEGIN");
    await client.query(
      "UPDATE accounts SET balance = balance - $1 WHERE id = $2",
      [amount, fromId],
    );
    await client.query(
      "UPDATE accounts SET balance = balance + $1 WHERE id = $2",
      [amount, toId],
    );
    await client.query("COMMIT");
  } catch (err) {
    await client.query("ROLLBACK");
    throw err;
  } finally {
    client.release();
  }
}

Why good: Dedicated client via pool.connect(), ROLLBACK on error, client.release() in finally guarantees the client returns to the pool

// ❌ Bad Example - Transaction with pool.query()
await pool.query("BEGIN");
await pool.query("UPDATE accounts SET balance = balance - 100 WHERE id = 1");
await pool.query("UPDATE accounts SET balance = balance + 100 WHERE id = 2");
await pool.query("COMMIT");
// Each pool.query() may use a DIFFERENT client -- the BEGIN/COMMIT
// execute on different connections, so there is no transaction at all

Why bad: pool.query() checks out a random client each time -- BEGIN, UPDATEs, and COMMIT may run on different connections, so there is no actual transaction

---

Pattern 4: Error Handling with PostgreSQL Error Codes

PostgreSQL errors include a code field with the SQLSTATE error code. See reference.md for the full error code table.

// ✅ Good Example - Handling specific PostgreSQL errors
const PG_UNIQUE_VIOLATION = "23505";
const PG_FOREIGN_KEY_VIOLATION = "23503";
const PG_DEADLOCK_DETECTED = "40P01";
const PG_SERIALIZATION_FAILURE = "40001";

interface PgError extends Error {
  code: string;
  constraint?: string;
  detail?: string;
  table?: string;
  column?: string;
}

function isPgError(err: unknown): err is PgError {
  return err instanceof Error && "code" in err;
}

try {
  await pool.query("INSERT INTO users (email) VALUES ($1)", [email]);
} catch (err) {
  if (isPgError(err) && err.code === PG_UNIQUE_VIOLATION) {
    throw new ConflictError(`Email already exists: ${err.constraint}`);
  }
  if (isPgError(err) && err.code === PG_DEADLOCK_DETECTED) {
    // Retry the operation
  }
  throw err;
}

Why good: Named constants for error codes (no magic strings), type guard for safe property access, specific handling per error type, re-throws unknown errors

---

Pattern 5: Streaming Large Result Sets

Use pg-cursor or pg-query-stream for queries returning many rows. See examples/streaming.md for full streaming patterns.

// ✅ Good Example - Cursor for batch processing
import Cursor from "pg-cursor";

const BATCH_SIZE = 100;

async function processAllOrders(pool: pg.Pool): Promise<void> {
  const client = await pool.connect();
  try {
    const cursor = client.query(
      new Cursor("SELECT * FROM orders WHERE status = $1", ["pending"]),
    );

    let rows = await cursor.read(BATCH_SIZE);
    while (rows.length > 0) {
      await processBatch(rows);
      rows = await cursor.read(BATCH_SIZE);
    }

    await cursor.close();
  } finally {
    client.release();
  }
}

Why good: Processes rows in fixed-size batches without loading entire result set into memory, proper client release in finally

---

Pattern 6: LISTEN/NOTIFY

PostgreSQL can push real-time notifications to connected clients. See examples/streaming.md for full examples.

// ✅ Good Example - LISTEN/NOTIFY with dedicated client
const CHANNEL = "order_updates";

async function listenForUpdates(pool: pg.Pool): Promise<pg.PoolClient> {
  const client = await pool.connect();

  client.on("notification", (msg) => {
    if (msg.channel === CHANNEL && msg.payload) {
      const data = JSON.parse(msg.payload);
      handleOrderUpdate(data);
    }
  });

  await client.query(`LISTEN ${CHANNEL}`);
  return client; // Caller is responsible for release on shutdown
}

// Publishing from another connection
await pool.query("SELECT pg_notify($1, $2)", [CHANNEL, JSON.stringify(data)]);

Why good: Dedicated client stays checked out for the lifetime of the listener, pg_notify() with parameterized channel/payload prevents injection, JSON payload for structured data

When to use: Real-time notifications where sub-second latency matters and the volume is low-to-moderate (hundreds per second). For high-throughput streaming, use a dedicated message broker.

---

<decision_framework>

Decision Framework

pool.query() vs pool.connect()

Do I need a dedicated client?
├─ Single query, no transaction? -> pool.query() (auto-releases)
├─ Multiple queries in a transaction? -> pool.connect() + BEGIN/COMMIT/ROLLBACK
├─ LISTEN for notifications? -> pool.connect() (keep client for lifetime of listener)
├─ Cursor/streaming? -> pool.connect() (cursor binds to a connection)
└─ Prepared statements across queries? -> pool.connect() (plan caches per connection)

Error Handling Strategy

What kind of PostgreSQL error?
├─ 23505 (unique_violation)? -> Map to 409 Conflict, include constraint name
├─ 23503 (foreign_key_violation)? -> Map to 400 Bad Request, entity not found
├─ 23502 (not_null_violation)? -> Map to 400 Bad Request, missing required field
├─ 23514 (check_violation)? -> Map to 400 Bad Request, validation failed
├─ 40P01 (deadlock_detected)? -> Retry with backoff (safe to retry)
├─ 40001 (serialization_failure)? -> Retry with backoff (safe to retry)
├─ 57014 (query_canceled)? -> Timeout, consider increasing statement_timeout
├─ 08xxx (connection_exception)? -> Pool handles reconnection, log and retry
└─ Other? -> Log full error, return 500

Streaming Decision

How many rows will the query return?
├─ < 1,000 rows? -> pool.query() is fine (result fits in memory)
├─ 1,000 - 100,000 rows? -> pg-cursor with batch processing
├─ 100,000+ rows? -> pg-query-stream piped to a writable stream
└─ Need to export to file? -> pg-query-stream piped to file write stream

</decision_framework>

---

<red_flags>

RED FLAGS

High Priority Issues:

• Using string interpolation/concatenation for SQL values -- this is SQL injection, the most dangerous vulnerability in database code
• Using pool.query() for transactions -- each call may use a different connection, so BEGIN/COMMIT have no effect
• Not releasing clients after pool.connect() -- leaked clients exhaust the pool; once max clients leak, the app deadlocks on pool.connect()
• Missing pool.on("error") handler -- idle client errors are emitted on the pool; unhandled, they crash the Node.js process
• Using standalone Client in application code -- no pooling, no reconnection, no concurrency

Medium Priority Issues:

• SELECT * in production queries -- returns unnecessary columns, breaks when schema changes, prevents index-only scans
• Loading millions of rows with pool.query() instead of streaming -- causes memory exhaustion and GC pressure
• Hardcoded connection strings -- prevents environment-specific configuration, risks credential leaks in version control
• Not handling specific PostgreSQL error codes -- generic error handling loses valuable information (which constraint, which column)
• Using LISTEN with pool.query() -- notifications bind to a specific connection; pool.query releases the connection immediately

Common Mistakes:

• Forgetting that result.rows[0] can be undefined when no rows match -- always check before accessing
• Relying on result.rowCount for SELECT emptiness checks -- use result.rows.length instead; rowCount is null for some commands (e.g., LOCK) and rows.length is universally reliable
• Using $1 inside string literals in SQL -- '$1' is a literal string, not a parameter; use $1 outside quotes
• Forgetting that PostgreSQL arrays in parameters are automatically converted -- [1, 2, 3] becomes {1,2,3} which works for = ANY($1) but not for IN ($1) (use = ANY($1::int[]) instead of IN)
• Calling client.release(true) routinely -- passing true destroys the client instead of returning it to the pool; only use after unrecoverable errors

Gotchas & Edge Cases:

• Pool error event vs query errors: Pool error fires for idle client backend disconnections (e.g., server restart). Query errors are thrown/rejected from the query call itself. You need both handlers.
• connectionTimeoutMillis: 0 (default) means no timeout -- connections wait forever if the pool is exhausted. Always set a timeout in production.
• idleTimeoutMillis only affects clients that have been returned to the pool -- a checked-out client that is never released will never be cleaned up.
• PostgreSQL numeric/decimal types are returned as strings by default (to avoid JavaScript floating-point precision loss). Parse them explicitly if you need numbers.
• LISTEN survives transactions -- if you BEGIN, LISTEN channel, ROLLBACK, the listener is still active. LISTEN is not transactional.
• pool.end() waits for all checked-out clients to be released. If a client is leaked (never released), pool.end() hangs forever.
• SSL connections: if the connection string contains any SSL parameters (sslmode, sslcert, sslkey, sslrootcert), the entire ssl config object is replaced -- use one or the other, not both.

</red_flags>

---

<critical_reminders>

CRITICAL REMINDERS

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST use parameterized queries ($1, $2, ...) for ALL values -- NEVER concatenate or interpolate user input into SQL strings)

(You MUST use Pool for all database access -- NEVER create standalone Client instances in application code)

(You MUST release clients back to the pool in a finally block after pool.connect() -- leaked clients exhaust the pool and hang the application)

(You MUST handle the error event on Pool instances -- unhandled idle client errors crash the Node.js process)

Failure to follow these rules will cause SQL injection vulnerabilities, connection pool exhaustion, application hangs, and process crashes.

</critical_reminders>