It's distributed, with no single point of failure. Your machines or vms that manage the cluster (should be) separate from those running the containers.

Ideally you set up 3 master nodes, with distributed etcd cluster for state, and enough machines to run your services with replication.

All of which is not bad at all to get started with if you use something like Kops.

Some useful docs: https://kubernetes.io/docs/admin/high-availability/building/