A team says their Kubernetes control plane is highly available because the API server runs three replicas. What is your assessment?

1. Locate the actual state: the API server, scheduler, and controller manager are stateless — every bit of cluster state lives in etcd. 2. Conclude that three stateless API servers in front of one etcd instance is not high availability, it is three doors to one single point of failure. 3. Specify etcd as a real cluster: an odd number of members, three or five, so a quorum survives a member loss, spread across failure domains. 4. Explain the quorum consequence explicitly: lose quorum and the cluster becomes read-only — no deploys, no scheduling, no scaling, though running pods keep serving. 5. Require scheduled snapshots stored outside the cluster and outside the same account, since a snapshot on the failing infrastructure is not a backup. 6. Insist on a rehearsed restore procedure, because etcd restore is order-sensitive and not something to learn during an outage. 7. Add the operational guardrails etcd specifically needs — low-latency disks, monitoring of member health and database size, and defragmentation. 8. Summarise: without a quorum etcd cluster and tested external snapshots, there is no disaster recovery plan, only a countdown.