Kubernetes Architecture — Complete Breakdown
etcd (pronounced et-see-dee) is the database that stores Kubernetes cluster state.
Think of Kubernetes as a distributed operating system for running containers across multiple computers.
Kubernetes has three major parts:
- Control Plane — the brain that makes decisions.
- Worker Nodes — the machines that actually run your applications.
- etcd — the database that remembers the cluster's state.
1. Kubernetes architecture diagram
Conceptual architecture of a Kubernetes cluster. The drawing simplifies communication paths and node components.
2. What exactly is etcd?
etcd Logo PNG Vector (AI, PDF, SVG) Free Download
etcd = Distributed Key-Value Database
It stores cluster configuration and state, such as Pods, Deployments, Services, Nodes and Secrets.
Imagine you tell Kubernetes:
I want three replicas of my Node.js application running.
Kubernetes must remember that you requested three replicas even if machines restart.
etcd stores the desired state and recorded cluster objects.
For example, conceptually:
/registry/deployments/default/my-app
desired replicas: 3
/registry/pods/default/my-app-xyz
node: worker-1
/registry/services/default/my-app
port: 80
These are illustrative representations, not the full underlying stored data.
Why etcd is important
- Persistent storage: Cluster state survives process restarts.
- Distributed consensus: Uses Raft to maintain agreement among etcd members.
- Consistency: Provides strongly consistent operations.
- High availability: Commonly deployed as a three- or five-member cluster.
- Watches: Allows the API server to observe changes efficiently.
The important detail is that Kubernetes components generally interact with etcd through the API server, rather than reading or modifying etcd directly.
3. All Kubernetes components explained
Control Plane — the brain
| Component | What it does |
|---|---|
| API Server | Entry point for Kubernetes. Validates requests and exposes the Kubernetes API. |
| etcd | Stores persistent cluster state. |
| Scheduler | Decides which node should run a new Pod. |
| Controller Manager | Continuously checks and corrects differences between desired and actual state. |
| Cloud Controller Manager | Connects Kubernetes with cloud infrastructure such as AWS or Google Cloud. Optional. |
Worker Node — the execution layer
| Component | What it does |
|---|---|
| Kubelet | Node agent responsible for making sure assigned Pods are running. |
| containerd | Common container runtime that manages containers and images. |
| runc | Low-level OCI runtime that creates containers using Linux features. |
| kube-proxy | Implements Kubernetes Service networking rules. Optional when replaced by another implementation. |
| CNI Plugin | Provides Pod networking. Examples: Cilium, Calico. |
| Pods | Smallest deployable Kubernetes units, containing one or more containers. |
These are the components described by the official Kubernetes architecture documentation.
Kubernetes+1
4. What happens when you deploy an application?
Suppose you execute:
kubectl create deployment my-app \
--image=nginx --replicas=3
- kubectl sends a request to the API Server
The request asks Kubernetes to create a Deployment with three replicas.
- API Server stores the Deployment in etcd
Kubernetes now remembers your desired configuration.
- Deployment and ReplicaSet controllers react
The controllers create the ReplicaSet and the three required Pod objects.
- Scheduler selects nodes
It checks resources and scheduling constraints, then assigns the Pods to suitable worker nodes.
- Kubelet starts the Pods
The kubelet on each selected node asks its container runtime to create the required containers.
- Container runtime launches the containers
containerd may use runc to create isolated Linux processes.
- Kubernetes continuously reconciles state
If a Pod disappears, the appropriate controller works to replace it and restore the desired replica count.
5. Interactive example: How reconciliation works
Change the desired number of replicas and simulate a Pod failure.
Deployment: my-app
Healthy
Desired state (etcd)
3 Pods
Actual state (running)
3 Pods
Desired replicas
3
Actual running Pods
Pod
Pod
Pod
Top comments (0)