Connection pooling

Every size has a max_connections ceiling (see the size catalogue). The engine enforces it and the gateway rejects the excess early with a protocol-appropriate error rather than letting clients hang.

Long-running services

Use your driver’s pool with a size well under the ceiling, divided by the number of replicas of your service:

Runtime	Setting
Node (pg, mysql2)	max / connectionLimit: 10 per process
Python (psycopg_pool, SQLAlchemy)	max_size / pool_size: 5 to 10, max_overflow: 0
Go (pgxpool, database/sql)	pool_max_conns / SetMaxOpenConns: 10
JVM (HikariCP)	maximumPoolSize: 10, connectionTimeout: 10 s
.NET (Npgsql, MySqlConnector)	Max Pool Size: 20 (per process)
Rails, Django, Laravel	Match the web-worker count; see the framework guides

Reserve roughly 10 connections for backups, migrations and psql.

Serverless and edge: the built-in pooler (PostgreSQL)

Functions open a connection per invocation and cannot share a process pool. PostgreSQL instances include a transaction-mode pooler (PgBouncer semantics) that multiplexes thousands of short client connections onto a small server pool.

• Enable it on Instance → Connect → Pooler. The pooled endpoint has the same host with -pool appended and the same port; the connection string is shown next to the direct one.
• Use the pooled endpoint from Vercel, Netlify, Cloudflare Workers, Lambda and Cloud Run; use the direct endpoint for migrations, LISTEN/NOTIFY, advisory locks and anything that needs session state.
• Prepared statements work per transaction; ORMs need the flag for that (?pgbouncer=true for Prisma, prepare: false for postgres.js, statement_cache_size=0 for asyncpg).

Not a substitute for a pool in long-running services

The pooler removes the per-connection cost of TLS and process forking; it does not raise the amount of work the instance can do. If your pooled endpoint waits on server connections, resize.

Other engines

• MySQL: no transaction pooler yet; use the driver pool and keep it small on serverless (2 to 5).
• Valkey: one multiplexed client per process; no pool needed.
• FerretDB: the MongoDB driver’s pool (maxPoolSize); each connection is a PostgreSQL connection underneath.
• libSQL: HTTP; no connection to keep. Use embedded replicas on runtimes with a filesystem.