Postgres Connection Pooling

TL;DR: Pick a pool mode based on what session-state your app uses, not on raw concurrency. Session mode is safe but caps you at one client per backend. Transaction mode multiplexes — but breaks LISTEN, SET, session advisory locks, WITH HOLD cursors, and historically prepared statements (fixed in pgBouncer 1.21+ with maxpreparedstatements). AWS RDS Proxy for Postgres pins on every SET, PREPARE, temp object, or 16+ KB statement. Always tune idleintransactionsessiontimeout to prevent leaks.

Jump to your fire

Decision diagram

flowchart TD
  A[App needs Postgres at high concurrency] --> B{App uses LISTEN/NOTIFY,<br/>session advisory locks,<br/>WITH HOLD cursors,<br/>or session-scoped SET?}
  B -->|Yes — heavy session state| C[Use SESSION mode]
  B -->|No, mostly autocommit + transactions| D{Using named prepared statements?<br/>Prisma, postgres-js, asyncpg, etc.}
  D -->|Yes| E{Pooler version?}
  E -->|pgBouncer ≥1.21 / Supavisor with PS support| F[Transaction mode<br/>+ max_prepared_statements ≥ 200]
  E -->|RDS Proxy / older pooler| G[Disable client-side prepared statements<br/>or use session mode]
  D -->|No| H[Transaction mode<br/>maximum multiplexing]
  C --> I[Tune idle_in_transaction_session_timeout]
  F --> I
  G --> I
  H --> I

1. Pool mode comparison

From pgbouncer.org/features.html, the three pool modes:

Feature compatibility (from pgbouncer.org/features.html)

¹ "You need to change maxpreparedstatements to a non-zero value to enable this support."

The decision rule: if your code uses any of the "Never" features, you must be in session mode (or refactor away the dependency). The most common silent killers are:

• ORMs that emit SET search_path = ... per session (turns transaction mode into pinned-session-mode)
• Background workers that hold session-level advisory locks
• Code using LISTEN for change notifications

2. Prepared statements: the pgBouncer 1.21+ flip

The historical advice — "disable prepared statements when using pgBouncer transaction mode" — flipped in late 2023. From the pgBouncer 1.21.0 changelog (2023-10-16, "The one with prepared statements"):

Add support for protocol-level named prepared statements! This is probably one of the most requested features for PgBouncer. Using prepared statements together with PgBouncer can reduce the CPU load on your system a lot.

pgBouncer 1.22+ enables it by default with maxpreparedstatements = 200.

From pgbouncer.org/config.html:

When this is set to a non-zero value PgBouncer tracks protocol-level named prepared statements... PgBouncer makes sure that any statement prepared by a client is available on the backing server connection. Even when the statement was originally prepared on another server connection.

The two caveats that still bite

1. SQL-level PREPARE/EXECUTE/DEALLOCATE are not tracked — only protocol-level (extended-query) prepared statements are. Code that does PREPARE mystmt AS SELECT ...; EXECUTE mystmt(...) still breaks in transaction mode.

2. The "cached plan must not change result type" gotcha:

If the return or argument types of a prepared statement changes across executions then PostgreSQL currently throws... ERROR: cached plan must not change result type... One of the most common ways of running into this issue is during a DDL migration where you add a new column or change a column type.

After a column-altering migration, every long-lived pooled backend may hold a stale plan. Two fixes: rolling-restart the application after migrations, or send DISCARD ALL (pgBouncer forwards this) to flush plans cluster-wide.

Client-side opt-outs (still useful for older poolers / RDS Proxy)

3. Supabase port-per-mode pattern

From supabase.com/docs/guides/database/connecting-to-postgres:

The session mode connection string connects to your Postgres instance via a proxy.

The transaction mode connection string connects to your Postgres instance via a proxy which serves as a connection pooler.

Critical rule: Supavisor and PgBouncer share the Postgres max_connections budget. From the docs:

Supavisor and PgBouncer work independently, but both reference the same pool size setting.

Supavisor backend connections + PgBouncer backend connections ≤ Postgres max_connections. Misconfigured pool sizes here are a common cause of Supabase outages on heavy load.

4. AWS RDS Proxy pinning triggers

From docs.aws.amazon.com/AmazonRDS/latest/UserGuide/rds-proxy-pinning.html:

When a connection is pinned, each later transaction uses the same underlying database connection until the session ends. Other client connections also can't reuse that database connection until the session ends.

A pinned session reverts to session-mode behavior — multiplexing is lost. Heavy pinning silently caps your effective concurrency.

Pinning triggers for RDS Proxy + PostgreSQL (from the docs)

• Any statement with text size > 16 KB (all engines)
• Using SET commands (no exceptions for Postgres — every SET pins)
• PREPARE, DISCARD, DEALLOCATE, or EXECUTE for prepared statements
• Creating temporary sequences, tables, or views

RDS Proxy doesn't support session pinning filters for PostgreSQL.

(MySQL has filter support; Postgres does not. So the only mitigation is not triggering pin in the first place.)

Mitigations

If you use SET statements to perform identical initialization for each client connection, you can do so while preserving transaction-level multiplexing. In this case, you move the statements that set up the initial session state into the initialization query used by a proxy.

Steps:

1. Move SET searchpath, SET timezone, SET applicationname to the proxy's InitQuery.
2. Disable client-side named prepared statements (see §2 table) for libraries that emit PREPARE.
3. Avoid CREATE TEMP TABLE patterns; use CTEs or transient real tables.
4. Watch the CloudWatch metric DatabaseConnectionsCurrentlySessionPinned — set an alarm if it grows past a small fraction of max_connections.

5. pgstatactivity diagnostics

From postgresql.org/docs/current/monitoring-stats.html:

The pgstatactivity view will have one row per server process, showing information related to the current activity of that process.

Key columns: pid, usename, applicationname, clientaddr, backendstart, xactstart, querystart, statechange, waiteventtype, state (active, idle, idle in transaction, idle in transaction (aborted)), query, backend_type.

The leak query

-- Find sessions stuck idle in transaction holding locks
SELECT
  pid,
  usename,
  application_name,
  client_addr,
  state,
  now() - state_change AS idle_duration,
  query
FROM pg_stat_activity
WHERE state IN ('idle in transaction', 'idle in transaction (aborted)')
  AND now() - state_change > interval '60 seconds'
ORDER BY state_change ASC;

The fix: timeouts

From postgresql.org/docs/current/runtime-config-client.html:

idleintransactionsessiontimeout — Terminate any session that has been idle (that is, waiting for a client query) within an open transaction for longer than the specified amount of time... A value of zero (the default) disables the timeout.

Even when no significant locks are held, an open transaction prevents vacuuming away recently-dead tuples that may be visible only to this transaction.

Recommended baseline: SET idleintransactionsessiontimeout = '60s' cluster-wide. Stops the most common cause of unbounded bloat.

idlesessiontimeout — Terminate any session that has been idle... but not within an open transaction... Be wary of enforcing this timeout on connections made through connection-pooling software or other middleware, as such a layer may not react well to unexpected connection closure.

Don't enable idlesessiontimeout if pgBouncer/Supavisor/RDS Proxy is in front; the pool will reconnect, but the disconnect storms can mask other problems.

Connection-string cheatsheet

Anti-patterns

Novice / Expert / Timeline

Timeline test: how long after a column-changing migration does a stale-cached-plan error reach an end user? Expert answer: zero (rolling restart in the deploy), or seconds (DISCARD ALL emitted from migration tooling). Novice answer: until the pgBouncer process restarts, hours later.

Quality gates

A pool change ships when:

• Test: Pool mode is explicit in config (not default-relying); CI lints for missing poolmode / maxprepared_statements.
• Test: idleintransactionsessiontimeout is set on the database; verified by a startup health check that runs SHOW idleintransactionsessiontimeout.
• Test: A leak-detector query (the §5 query) runs as a Grafana alert; threshold = "any session > 60s idle in transaction".
• Test (RDS Proxy): CloudWatch alarm on DatabaseConnectionsCurrentlySessionPinned > N.
• Test (transaction mode): Integration test that runs the app's full smoke suite against a pgBouncer in transaction mode — catches LISTEN, session advisory lock, etc. usage that'd silently break in prod.
• Manual: Migration runbook includes either rolling restart or DISCARD ALL to flush cached plans across pooled backends.

NOT for this skill

• ORM-specific connection config (use prisma-connection-config, sqlalchemy-async-engine, etc.)
• Postgres replication / read replicas (use postgres-replication-design)
• Non-Postgres pooling (PgPool, ProxySQL for MySQL) — different feature sets
• General Postgres performance tuning (use postgres-explain-analyzer)
• Long-running migrations themselves (use zero-downtime-database-migration)

Sources

• pgBouncer: Features and Pool Modes
• pgBouncer: Configuration reference — maxpreparedstatements, pool_mode, etc.
• pgBouncer: Changelog 1.21.0 — "The one with prepared statements"
• Supabase: Connecting to your database — direct vs Supavisor session vs transaction
• AWS: RDS Proxy pinning
• PostgreSQL: Monitoring statistics — pgstatactivity
• PostgreSQL: Client connection defaults — idleintransactionsessiontimeout