Connection pool tuner

Connection Pool Tuner

Pool exhaustion shows up as checkout timeouts and "too many clients" errors; the fix is almost never "raise max_connections." Diagnose from evidence first, then size to the database.

Workflow

1. Capture the evidence before touching any number. Query pg_stat_activity grouped by state under load. Record total connections, count per state, and the longest-running query/transaction age. Also record every pool that points at the DB: app replicas times per-replica pool size, plus migrations and any superuser sessions.

2. Classify the failure from the evidence:

   • Many idle in transaction rows -> code opens a transaction and does slow work (an HTTP call, sleep, batch) before commit. Fix the transaction scope, not the pool.
   • Many active long-runners -> slow queries saturating connections. Tune the queries.
   • Checkout/acquire timeouts while DB connections sit mostly idle -> the app pool is too small or connections leak (not returned to the pool). Find the leak before enlarging.
   • Total connections at or above max_connections -> the sum of pools is too large; reduce per-replica pool or front with a proxy. Do not raise max_connections.

3. Size the pool to the database, not the app. Postgres performs best at a small connection count; each backend is a process with real memory cost. Target tens of active connections, not thousands; a common ceiling is roughly (core_count * 2) + effective_spindle_count. The sum of all pools must stay under max_connections with headroom for migrations and superuser - keep at least 10-15% of max_connections free. Twenty replicas with a pool of 20 each is 400 connections - usually far too many.

4. Set timeouts that fail fast. Set a pool checkout/acquire timeout (2-5s) so a request errors instead of piling up. Set idle_in_transaction_session_timeout (30-60s is a sane starting range for web workloads) to kill stuck transactions. Set statement_timeout to cap runaway queries - anchor it to your slowest legitimate query plus headroom, often 30s for OLTP paths. In HikariCP, set connectionTimeout and keep maxLifetime a minute or two below the DB/proxy idle cutoff so you never hand out a dead socket.

5. Multiplex with PgBouncer when connection count is the bottleneck. Run PgBouncer in transaction pooling mode to fold many client connections onto few server connections. Transaction mode forbids session-level features - prepared-statement caches, SET, advisory locks held across statements - so configure clients accordingly.

6. Tame serverless storms. Serverless/Lambda functions each open their own connection and storm the DB on a spike. Route them through PgBouncer or a serverless pooler/driver (Neon or Supabase pooler, RDS Proxy, Prisma Accelerate) and keep per-function pool size at 1.

7. Verify under load. Re-run the pg_stat_activity query and confirm the failing state is gone and total connections sit safely under max_connections.

Deliverable

Produce a tuning record alongside the config change: the before/after pg_stat_activity readings, the failure classification from step 2 with the evidence that supports it, the full connection budget (every pool times its replicas, plus migration/superuser headroom, summed against max_connections), and each changed setting - pool size, checkout timeout, idle_in_transaction_session_timeout, statement_timeout, maxLifetime, PgBouncer mode - with its one-line justification. The record proves the change fixed the diagnosed cause rather than masking it.

Quality bar

• Every size or timeout change is justified by a captured pg_stat_activity reading, not a guess.
• The sum of all pools, plus migration and superuser headroom, stays under max_connections.
• Each failure mode is traced to its real cause (transaction scope, slow query, or leak) before any pool is enlarged.

Do NOT

• Do NOT raise max_connections or enlarge the pool as the first move - that masks a leak or a slow transaction and adds per-backend memory pressure.
• Do NOT prescribe numbers without a captured pg_stat_activity reading.
• Do NOT budget pools per replica in isolation - the database sees the sum across every replica, sidecar, and cron job, and that sum is what exhausts max_connections.
• Do NOT use transaction-mode PgBouncer with clients that rely on session prepared statements, SET, or cross-statement advisory locks unless reconfigured.
• Do NOT leave per-function serverless pools larger than 1.
• Do NOT use this skill for external-service flakiness (retries, fallbacks, backoff) - that is circuit-breaker-builder's job.

Defaults to reach for

Start small: an app pool of 5-10 per replica, a checkout timeout of 2-5s, and a proxy in transaction mode if connection count is the bottleneck. If raising the pool fixes latency, you had a leak or a slow transaction - find it rather than masking it.