Performance
A Slow App Is Often Just a Database Waking Up
Free tiers suspend on idle. The first request after a pause pays a cold start, and it looks exactly like a slow product.
A user reports that your web application took six seconds to load. You open the site in your browser: 180 milliseconds. You check the server logs: median latency is 95ms. You close the bug report as 'cannot reproduce'.
The bug was real. The user was hitting your serverless database cold start.
The idle suspension penalty
Modern serverless architectures (Neon, Supabase free compute, Railway starter tiers, AWS Lambda) aggressively scale down idle compute to zero. When a product is in its early stages or has regional traffic dips, the first request after thirty minutes of silence pays the wake-up tax.
- Container spin-up: 800ms - 1,500ms.
- TLS handshake to database proxy: 200ms - 400ms.
- Postgres buffer pool warming: 500ms - 1,000ms.
- Initial query execution: 50ms.
To the user, your product felt like an abandoned, broken prototype. To your monitoring tool, which measures average latency across warm requests, everything looked green.
How to eliminate the wake-up gap
You do not need an enterprise tier to fix this. Two simple patterns solve it:
- Scheduled synthetic keep-alives: a light cron job pinging a `/health` endpoint that runs a trivial query every four minutes keeps the connection pool and compute warm.
- Prerender static shells: deliver the navigation, layout, and cached read-only content instantly from static storage while background data streams in via React Suspense or asynchronous fetches.
A fast first paint buys you five seconds of user patience. A white blank screen buys you zero.