Stop 08 · Kubernetes & OrchestrationTopics · 1/12

Kubernetes Architecture

In short

Kubernetes looks like a hundred moving parts until you see the split: a small control plane that decides, and worker nodes that do. Every other concept hangs off that one idea.

A control plane (API server, scheduler, controller manager, etcd) manages a cluster of worker nodes running your containerized workloads — understanding this split clarifies almost everything else about Kubernetes.

The mascot pulls levers in a railway control room while toy trains run on parallel tracks.

Think of it like this

The railway control room

Trains don't decide their own routes. A control room watches every track, sets signals and sends a spare engine when one breaks down. The drivers just follow instructions. The control plane is that control room; worker nodes are the trains doing the actual running.

Watch it happen

From kubectl apply to a running pod

kubectl

API server

Scheduler

kubelet

apply deploy.yaml

1/6You never talk to nodes directly. Every request goes to the API server, the front desk of the cluster.

The big ideas

What you need to know

Control plane decides

API server, etcd, scheduler and controller manager. They store what you want and keep working to make it true.

Nodes do the work

Each worker runs a kubelet, a container runtime and kube-proxy. Your pods live here.

Desired vs actual

You declare 'three copies'. Controllers loop forever comparing that with reality and fixing any gap.

etcd is the memory

All cluster state lives in etcd, a key-value store. Lose it without a backup and the cluster forgets everything.

See it in code

Walk through the code

bash

1# Start a local cluster (needs Docker)2kind create cluster --name lab3 4# Who's in the cluster?5kubectl get nodes -o wide6 7# Control plane parts run as pods in kube-system8kubectl get pods -n kube-system

1/3kind creates a one-node cluster where the same node runs both the control plane and your workloads.

kind runs a whole cluster inside Docker containers, free on your laptop. minikube and k3d are good alternatives.

Save yourself the pain

Common mistakes

Oops

Thinking pods are 'sent' to nodes by the API server.

Do this instead

Kubelets watch the API server and pull work assigned to them. Everything in Kubernetes is driven by watching and reconciling.

Oops

Running production workloads on control plane nodes to save money.

Do this instead

Keep the control plane separate, or use a managed service like EKS, GKE or AKS that runs it for you.

Your mission

Meet your control plane

~25 min

Quick check

You delete a pod that belongs to a Deployment. Which part notices and creates a replacement?

Go deeper

1Diagram the control plane and worker node components and what each does.
2Learn the role of etcd as the cluster's source of truth (a key-value store of all cluster state).
3Set up a local single-node cluster (minikube or kind) to explore this hands-on.