Neon

Postgres was designed for a server that stays up. Neon separates storage from compute and rebuilds the developer experience around that: the compute sleeps when nobody’s querying, wakes in well under a second, and branches — a full copy-on-write clone of your data, created instantly and costing almost nothing until you change something.

It’s still ordinary Postgres. Your driver, your migrations, your extensions — pgvector included — all work unchanged. Databricks acquired Neon in 2025; the service kept its shape and got cheaper.

Connect link

  psql "postgresql://user:pass@ep-cool-name-123456.us-east-2.aws.neon.tech/neondb?sslmode=require"
  

Two connection strings are offered and the difference matters:

• The pooled one (-pooler in the host) for anything running per request — serverless functions, edge handlers. Postgres charges memory per connection and serverless opens a lot of them.
• The direct one for migrations, which need a session that pooling won’t give them.

For Cloudflare Workers and other runtimes with no raw TCP, there’s an HTTP driver:

  npm install @neondatabase/serverless
  

  import { neon } from '@neondatabase/serverless';
const sql = neon(process.env.DATABASE_URL);
const rows = await sql`SELECT id, title FROM posts WHERE published = true`;
  

Branching is the reason to be here link

  npm install -g neonctl
neonctl branches create --name pr-482 --parent main
neonctl connection-string pr-482
neonctl branches delete pr-482
  

A branch is a copy-on-write clone of the parent at a point in time. Created in about a second, whatever the size of the database, and billed only for what diverges. Three things this makes routine:

• A database per pull request. CI creates a branch from production-shaped data, runs migrations and tests against it, and deletes it when the PR closes. Your integration tests stop running against an empty schema.
• Rehearsing a migration on real data before it touches production, at no meaningful cost.
• Point-in-time recovery as a branch. Someone ran the wrong UPDATE? Branch from twenty minutes ago, read the rows as they were, and repair — without restoring over anything.

The free tier, and the trade link

The free plan gives a project with a small storage allowance and a monthly compute budget, with scale-to-zero on by default. Read the two consequences honestly:

• Scale to zero means a cold start. Sub-second, but the first request after idle is slower. For a hobby app that’s fine; for a checkout flow, keep the compute pinned on a paid plan.
• Compute is metered. An always-busy app on the free tier will exhaust the budget before it exhausts storage.

Neon or Supabase? link

Both offer free Postgres, and they’re aimed at different things.

Choosing Neon means picking your own auth and storage — Clerk and R2 are the usual companions on this site.

Next link

The other end of the trade: a whole backend in one project → Supabase