Sleeping and waking

A database with nothing connected to it stops after five minutes and starts again on the next connection.

This is the single fact that makes a database per project affordable, and the single thing that will surprise you.

What it costs you

The first connection after a quiet spell waits a second or two while the compute starts. Everything after that is a normal Postgres connection with no penalty at all, and the compute stays up for five minutes past the last one.

An application with steady traffic never sees it. A staging environment nobody touched since Friday sees it on Monday morning, once.

What it saves you

Compute is billed only while it is awake. A project that gets a visitor an hour pays for a few minutes of compute a day rather than for a server running all night — which is the difference between a database costing pennies and costing dollars.

Storage is unaffected: your data is held whether the compute is up or down, and that is the part the minimum covers.

Do not keep it awake

The obvious trick — a cron job pinging the database every four minutes — costs more than it saves. It converts a few minutes of daily compute into twenty-four hours of it, to avoid a delay measured in seconds, on requests nobody is waiting for.

If a cold start genuinely hurts on a user-facing path, the honest fixes are:

• Connect at the edge of the request, not lazily mid-render. The wake happens in parallel with the rest of your page rather than after it.
• Cache what does not need to be live. A homepage that reads the same six rows every time should not be waking a database to do it.
• Tell us. The idle timeout is per project and adjustable; longer costs more and we will say by how much.

Autoscaling, while it is awake

The compute moves between 0.25 and 4 CU on its own, following what your queries need. You are billed for what it actually used, per compute-hour, not for the ceiling.

Neon's own default for a new project is a fixed 1 CU. Ours starts at a quarter of that deliberately: a quiet site should not pay for a whole compute unit because somebody visited it.

If your workload needs more than four CU, that is a conversation rather than a limit — but it is usually a query that wants an index.