Draft: not yet run against a live cluster. Treat output as representative and check it on yours.
When the control plane fails, running workloads usually keep running. What stops is change: no deploys, no scaling, no self-healing. That makes control plane incidents deceptive. Everything looks fine until something needs to move.
Everything goes through the API server. No component talks to etcd except the API server. If it’s unreachable, the cluster can’t change.
etcd needs a quorum. With three members, the cluster survives one failure; with five, two. Lose quorum and writes stop.
The scheduler and controller manager run with leader election. Only one replica of each is active, holding a Lease in kube-system. If the leader dies, another takes over after the lease expires.
Webhooks sit in the request path. Mutating and validating admission webhooks are called synchronously for matching requests. A slow or dead webhook makes those requests slow or failed, including requests from controllers trying to replace pods.
On a kubeadm or kind cluster, the control plane components are static pods on the control plane node. On kind, you can stop one to see the effect: docker exec <cluster>-control-plane mv /etc/kubernetes/manifests/kube-scheduler.yaml /tmp/, create a pod and watch it stay Pending with no events, then move the file back.
The API server is down. What happens to running pods?
They keep running. The kubelet keeps existing containers alive, but nothing can be created, scaled, or rescheduled until the API server is back.
New pods stay Pending with no events at all. Which component do you suspect?
The scheduler. Without it, nothing records FailedScheduling or binds the pod.
Why can a validating webhook with failurePolicy Fail make a cluster unable to recover?
Every matching request is rejected while the webhook is down, including requests to create the webhook’s own replacement pods if it matches its own namespace.
Maps to: CKA: Troubleshooting; CKA: Cluster Architecture, Installation and Configuration. Check the current exam curriculum, since domain names change between revisions.