上QQ阅读APP看书，第一时间看更新

etcd cluster

To persist the cluster state, Kubernetes uses etcd – a distributed, reliable key-value store that utilizes the Raft distributed consensus algorithm in order to provide sequential consistency. The etcd cluster is the most important part of the control plane – this is the source of truth for the whole cluster, both for the current state and the desired state of the cluster.

Generally, single-node etcd clusters are only recommended for testing purposes. For production scenarios, you should always consider running at least a five-member cluster (with an odd number of members) in order to provide sufficient fault tolerance.

When choosing an etcd cluster deployment topology, you can consider either a stacked etcd topology or an external etcd topology. A stacked etcd topology consists of one etcd member per Kubernetes master instance, whereas an external etcd topology utilizes an etcd cluster deployed separately from Kubernetes and is available via a load balancer. You can find out more about these topologies in the official documentation: https://kubernetes.io/docs/setup/production-environment/tools/kubeadm/ha-topology/.

The watch protocol that's exposed by etcd is also a core functionality for the watch API in Kubernetes, which is provided by kube-apiserver for other components.