Describe the Kubernetes architecture: control plane vs worker nodes and their components.

Question

IT Interview Practice · Accepted Answer

A Kubernetes cluster has two kinds of machines: the **control plane** (the "brain" that makes decisions and stores state) and **worker nodes** (where your Pods actually run). Each has a small set of well-defined components.

## The big picture

```text
┌──────────────── CONTROL PLANE (the brain) ────────────────┐
│                                                            │
│   kube-apiserver ── the front door; ALL comms go through it│
│        │                                                   │
│   etcd ───────── key-value store: the cluster's SOURCE OF  │
│                  TRUTH (all objects & desired state)       │
│   kube-scheduler ── decides WHICH node a new Pod runs on   │
│   controller-manager ── runs control loops (reconcilers)   │
└──────────────────────────┬─────────────────────────────────┘
                           │ (apiserver)
   ┌───────────────────────┼───────────────────────┐
┌──┴─── WORKER NODE ──┐ ┌──┴─── WORKER NODE ──┐  ...
│ kubelet             │ │ kubelet             │
│ kube-proxy          │ │ kube-proxy          │
│ container runtime   │ │ container runtime   │
│  └ Pods (your apps) │ │  └ Pods (your apps) │
└─────────────────────┘ └─────────────────────┘
```

## Control plane components

```text
kube-apiserver        The single entry point. Exposes the REST API; validates &
                      persists objects to etcd. kubectl, controllers, and kubelets
                      all talk ONLY to the apiserver.
etcd                  Distributed, consistent key-value store. Holds the ENTIRE
                      cluster state (desired + current). Lose etcd = lose the cluster.
kube-scheduler        Watches for Pods with no node assigned and picks the best node
                      (based on resource requests, affinity, taints, etc.).
kube-controller-      Runs the control loops: node controller, replication (ReplicaSet),
manager               deployment, job, endpoints, etc. Each loop drives actual → desired.
cloud-controller-     (On cloud) integrates with the provider: load balancers, routes,
manager               node lifecycle.
```

## Worker node components

```text
kubelet          The node agent. Talks to the apiserver, starts/stops containers via
                 the runtime, and reports Pod & node status. It makes the node's
                 reality match the Pods assigned to it.
kube-proxy       Programs the node's networking (iptables/IPVS) so that Service virtual
                 IPs are load-balanced to the right Pod endpoints.
container runtime The software that actually runs containers (containerd, CRI-O).
```

## How a "kubectl apply" flows through the system

```text
1. kubectl → kube-apiserver: "I want a Deployment with 3 replicas"
2. apiserver validates & writes the object to etcd
3. controller-manager (deployment→replicaset controller) sees it, creates 3 Pod objects
4. scheduler sees 3 unassigned Pods, assigns each to a node (writes back via apiserver)
5. each node's kubelet sees its Pods, tells the runtime to pull images & start containers
6. kube-proxy wires up Service networking so traffic reaches the new Pods
Everything is a WATCH + RECONCILE loop mediated by the apiserver.
```

```bash
kubectl get nodes                 # list nodes (control-plane + workers)
kubectl get pods -n kube-system   # see the control-plane / system components
kubectl get componentstatuses     # health of core control-plane pieces
```

## Why it matters

Understanding the split between the control plane and worker nodes — and the role of each component — is what turns Kubernetes from a black box into a system you can reason about and operate, so it is core knowledge for any middle-or-above engineer. The **control plane** is the decision-making brain: the **kube-apiserver** is the single front door through which every component and user communicates; **etcd** is the consistent key-value store that holds the entire cluster state and is the source of truth (losing it means losing the cluster); the **scheduler** decides which node each new Pod lands on; and the **controller-manager** runs the reconciliation loops that continuously drive actual state toward desired state. On each **worker node**, the **kubelet** is the agent that makes the node's containers match the Pods assigned to it, **kube-proxy** programs the networking that makes Service virtual IPs route to the right Pods, and a **container runtime** (containerd, CRI-O) actually runs the containers. Seeing how a single `kubectl apply` flows through this system — apiserver to etcd, controllers creating objects, scheduler placing Pods, kubelets starting containers, kube-proxy wiring networking, all as watch-and-reconcile loops — reveals the elegant, consistent design behind the platform. This mental model is exactly what you need to debug why a Pod is stuck Pending (scheduler), why the cluster is unresponsive (apiserver/etcd), or why a Service isn't routing (kube-proxy), making it indispensable for operating and troubleshooting real clusters.

The big picture

Control plane components

Worker node components

How a "kubectl apply" flows through the system

Why it matters

Related questions