| 1 | The control plane is five components arranged around one hub. At the center sits kube-apiserver, the only component allowed to talk to etcd, the persistent key-value store on its left. Around the apiserver three reconciliation loops listen and act: the scheduler picks a node for every unscheduled Pod, the controller-manager runs the workload controllers, and the optional cloud-controller-manager translates cluster intent into cloud-provider primitives like load balancers and routes. Notice that every spoke watches the apiserver — none of them call each other directly. |