Shipping an AI SaaS on Cloudflare Workers

Running the whole stack — marketing site, dashboard, admin, and API — on Cloudflare Workers sounds aggressive. In practice the constraints are real but few, and they push you toward a cleaner architecture.

Create clients per request

There is no long-lived process to hold a connection pool. Every database client is created inside the request, over Neon's HTTP driver:

const db = createDb(env.DATABASE_URL);
const products = await db.select().from(schema.product);

No module-level I/O, no warm-pool assumptions — which also keeps cold starts small.

No interactive transactions

Neon over HTTP can't hold a transaction open across round-trips. For multi-step atomicity you reach for db.batch([...]), which Neon runs as a single transaction, or a single guarded statement.

What surprised us

• TTFB dropped under 100ms globally without touching a single config.
• verbatimModuleSyntax had to stay off in the web app, or server bundles leak into the client.
• Plugin order in vite.config.ts is load-bearing.

The payoff: one mental model, one deploy target, and latency that's fast everywhere instead of fast near one region.