Stop a rolling deploy from cutting off requests mid-flight

Every deploy, a handful of users get a connection reset in the middle of a request, and it always clears up right after.

IntermediateRead 5 minDoing: an afternoon7 steps

▶ Start

What you'll have at the end

A shutdown handler that stops taking new connections and finishes the requests already in progress before the process exits, proven clean against a live load test run through a real rollout, more than once.

Before you start

You need

A FastAPI service already running behind a Kubernetes Deployment and Service, with more than one replica so a rolling update actually replaces pods one at a time while the others keep serving.

Not covered

Sizing the readiness probe itself, so a brand new pod does not take traffic before it can actually handle it.

Leans on

7 steps

Where to go next

A health check lying about being ready

a brand new pod taking traffic before it can actually handle it is a readiness problem on the other end of the same rollout.

Checked 19 Aug 2026

Part of the Backend Deployment cookbook