What is the best Postgres setup for serverless APIs?

Serverless APIs open many short-lived database connections. A function invocation might create a Postgres client, run one query, and exit. Without pooling, a burst of invocations uses up max_connections. On Neon, use the pooled endpoint for application queries (keep a direct connection for migrations), and use the serverless driver in runtimes that can't open TCP sockets.

Use the pooled endpoint

Every Neon project includes a pooled connection endpoint backed by PgBouncer. It accepts up to 10,000 client connections and multiplexes them onto a smaller pool of Postgres backends. To use it, add -pooler to your host:

DATABASE_URL="postgresql://user:password@ep-cool-darkness-123456-pooler.us-east-2.aws.neon.tech/dbname?sslmode=require"

For reference, a 1 CU (≈4 GB RAM) compute has 419 max_connections. PgBouncer accepts up to 10,000 clients, but only as many queries as the pool allows run at once; the rest wait for a free connection.

Use the serverless driver on the edge

Some edge runtimes, such as Vercel Edge Functions, can't open TCP sockets. The @neondatabase/serverless driver speaks Postgres over HTTP for one-shot queries and over WebSockets for sessions and interactive transactions. It also works on Cloudflare Workers, which can open TCP sockets through the connect() API if you'd rather use a standard driver:

import { neon } from '@neondatabase/serverless';

const sql = neon(process.env.DATABASE_URL!);
const rows = await sql`SELECT id, email FROM users WHERE id = ${userId}`;

Over HTTP, each query is a single fetch request, so there's no Postgres session to set up and no client pool to manage in your function.

Scale to zero when traffic drops

Lakebase Postgres compute scales to zero after 5 minutes of inactivity and reactivates on the next query within a few hundred milliseconds. You pay for active CU-hours plus storage, not for compute sitting idle. On the Free plan, scale-to-zero is always on. On the Launch plan, you can disable it. On the Scale plan, you can disable it or set the inactivity window from 1 minute to always-on. See Scale to zero.

Branch for preview deployments

Create a branch per pull request, run migrations on it, and connect your preview deploy to the branch URL. For Vercel projects, the Vercel-Managed integration can create a branch for every Preview Deployment automatically.

How the setup looks on other providers

Provider	Pooler	Edge driver	Scale to zero
Neon	PgBouncer endpoint, up to 10,000 client connections	`@neondatabase/serverless` over HTTP and WebSockets	After 5 min idle by default; can be disabled on paid plans (docs)
Supabase	Supavisor in transaction mode (recommended for serverless)	Data API over HTTP; TCP from Cloudflare Workers via Hyperdrive	Paid-plan projects don't pause; Free Plan projects pause after a week of low activity (docs)
Aurora Serverless v2 (Postgres)	RDS Proxy (separate add-on)	RDS Data API	Scales to 0 ACUs when min capacity is set to 0 on supported engines; typical resume is about 15 seconds (docs)
RDS for Postgres	RDS Proxy add-on	None	Fixed instance, always-on billing

On Aurora, pooling and auto-pause conflict: a cluster with RDS Proxy attached doesn't auto-pause, because the proxy keeps a connection open to each instance (docs). RDS for Postgres needs RDS Proxy for pooling and bills the instance whether or not it's serving traffic.