Shipping a new version is a question of how many users meet it, and how fast you can take it back. The simplest answer is a rolling update. Suppose the app runs on 4 pods behind a load balancer. The platform starts one new pod, waits until it is healthy, then stops one old pod, and repeats.
After the first swap, 3 pods run the old version and 1 runs the new one, so about 25% of requests reach the new code. At no point do fewer than 3 of 4 pods serve traffic. The cost is a mixed period where both versions run side by side. For pipeline basics, see the System Design lesson "CI/CD & Deployments".
💡Good to know
Rolling needs no extra capacity, but old and new code must work together for a while.