Cluster

Machines are split into two groups:

• Control Plane - the brain of the cluster
• Workload Plane

Control Plane

It consists of:

• Kubernetes API Server - RESTful API used by users and other components to create objects
• Scheduler - decides which node runs which app
• Controllers - have different tasks. Most of them create objects from the objects users create (e.g., if I create a deployment, it will create pods) Some communicate with external systems (e.g. cloud provider API). Most object types have an associated controller (e.g. “Deployment Controller” manages “Deployment” objects).
• etcd - the K8s API Server stores objects here. It’s a distributed data store.

Workload Plane

It consists of:

• Kubelet - an agent that talks to K8s API Server and manages apps running on a node. It reports apps’ status via the API
• Container Runtime - runs containers
• Kube Proxy - load-balances traffic between apps.
• apps

There are also some addons that could be on Worker or Master nodes.

Failures

In case of a failure of a node, its status changes to NotReady. The pods that were supposed on that node are still kept as Running for the next few minutes. After that time, the pods will be recreated on other nodes. The exact timings can be configured for the cluster.