Architecture

A Nimbus deployment has three moving parts:

• Edge router — accepts every incoming request and routes it to the nearest region running your function.
• Function runtime — a lightweight V8 isolate (or, for Python, a sandboxed interpreter) that executes your code with a cold start under 5ms.
• Control plane — the service behind nimbus deploy, responsible for distributing new versions to all 30 regions and rolling back failed deploys.

Request lifecycle

1. A request hits the edge router in the region closest to the client.
2. The router checks for a warm isolate running the latest deployed version of your function. If none exists, it cold-starts one.
3. Your function runs and returns a Response.
4. The router streams that response back to the client and records a trace.

Note

Because each region runs its own copy of your function, there’s no single point of failure — a region outage only affects traffic routed to that region.

Deploys are atomic

nimbus deploy uploads your function to a staging slot in every region, health-checks it, then flips traffic over all at once. If any region fails its health check, the whole deploy is aborted and traffic stays on the previous version.