Deployment = manager, ReplicaSet = shift supervisor per version, Pods = riders.

A Deployment does not own Pods directly. Each revision of the Pod template gets its own ReplicaSet, whose single job is holding a Pod count. A rollout shifts replicas from the old ReplicaSet to the new one, and leaving the old one around at zero is what makes undo fast.

Read it once, then a question tests whether the trick stuck.