Architecture

Architecture

Kubernetes clusters consist of nodes

The cluster is maintained by the control plane

Applications are executed on worker nodes

Demo: Cluster nodes

$ kubectl get nodes
NAME                 STATUS   ROLES    AGE   VERSION
kind-control-plane   Ready    master   78s   v1.18.2
kind-worker          Ready    <none>   41s   v1.18.2

Worker Nodes 1/2

Worker nodes execute containerized services

kubelet manages pod descriptions and controls the container runtime

The container runtime manages the individual containers

kube-proxy is responsible for network communication with services

Worker Nodes 2/2

Demo: Process tree with running services

$ docker exec -it kind-worker ps fxwwo pid,cmd
  PID CMD
  123 /usr/local/bin/containerd
  430 /usr/bin/kubelet
  479 /usr/local/bin/containerd-shim-runc-v2
  528  \_ /pause
  598  \_ /bin/kindnetd
  504 /usr/local/bin/containerd-shim-runc-v2
  535  \_ /pause
  607  \_ /usr/local/bin/kube-proxy

Control plane 1/2

Control plane nodes manage the cluster

api-server provides the Kubernetes API

All cluster data is stored in etcd

etcd is a key-value store

kube-scheduler assigns pods to workers

kube-controller-manager runs controllers that adapt the cluster state to the resources

Control plane 2/2

Demo: Process tree with running services

$ docker exec -it kind-control-plane ps fxwwo pid,cmd
  PID CMD
  129 /usr/local/bin/containerd
  353 /usr/local/bin/containerd-shim-runc-v2
  472  \_ /pause
  660  \_ etcd
  357 /usr/local/bin/containerd-shim-runc-v2
  474  \_ /pause
  564  \_ kube-scheduler
  363 /usr/local/bin/containerd-shim-runc-v2
  459  \_ /pause
  571  \_ kube-controller-manager
  389 /usr/local/bin/containerd-shim-runc-v2
  458  \_ /pause
  620  \_ kube-apiserver

Controller

Controller are responsible for a specific resource type

Controllers are implemented as a infinite loop…

…that checks the current state…

…and tries to reach the desired state

Kubernetes ships with many builtin resource types, e.g. Pod, Service, Deployment

kube-controller-manager contains all builtin controllers

Example

1. A resource of type Pod is created
2. The API creates an event
3. kube-controller-manager reacts to the event
4. …and creates a pod via the API

Communication

From workers to the control plane

All services on the worker nodes communicate with the api-server

From the control plane to workers

kubelet handles all communication for the worker node

Cluster scale

Clusters can be scaled in different ways

etcd can run as a container in the cluster

etcd can be consumed as an external service