Building reliable and highly available distributed systems is a non-trivial endeavor. In this section, we will check some of the best practices that enable a Kubernetes-based system to function reliably and be available in the face of various failure categories.
To create a highly available Kubernetes cluster, the master components must be redundant. That means etcd must be deployed as a cluster (typically across three or five nodes) and the Kubernetes API server must be redundant. Auxiliary cluster-management services such as Heapster's storage may be deployed redundantly too, if necessary. The following diagram depicts a typical reliable and highly available Kubernetes cluster. There are several load-balanced master nodes, each one containing whole master components as well as an etcd component:
This is not the only way to configure highly available clusters. You may prefer, for example, to deploy a standalone etcd cluster...