Learning Kubernetes: Why, What, How, and the Theory Behind It

A structured guide for building real understanding of Kubernetes (K8s) — not just memorizing commands, but understanding the problems it solves and the ideas it's built on.

Before learning any tool, understand what pain it removes.

• Applications used to run on a few big servers. Scaling meant buying bigger hardware (“scale up”).
• Virtualization (VMs) let you split one machine into many, but VMs are heavy — each carries a full OS.
• Containers (via Docker, ~2013) made this lighter: package an app + its dependencies into one portable unit that runs the same everywhere.

Once you have containers, new questions appear:

• If a container crashes, who restarts it?
• If traffic spikes, who creates more copies (and removes them later)?
• If a physical server dies, who moves the containers elsewhere?
• How do hundreds of containers find and talk to each other?
• How do you roll out a new version without downtime?
• How do you manage secrets, configuration, and storage consistently across machines?

Doing this manually across a fleet of servers doesn't scale. You need a system that treats a cluster of machines as one programmable unit, and constantly works to keep your application in the state you asked for.

• Declarative model: you describe the desired end state (“I want 3 copies of this app running”), not the steps to get there. Kubernetes continuously works to make reality match your declaration.
• Self-healing: it detects failures and corrects them automatically.
• Portability: it runs the same way on AWS, GCP, Azure, bare metal, or your laptop — avoiding vendor lock-in.
• Ecosystem: it became the industry standard (born from Google's internal system, Borg), so most cloud-native tooling assumes Kubernetes underneath.

Core idea to hold onto: Kubernetes is an orchestrator. Docker/containers give you the packaging; Kubernetes gives you the management of many containers across many machines.

This is the conceptual model. Understanding this deeply matters more than memorizing YAML syntax.

Kubernetes is built on control theory — the same principle behind a thermostat.

1. You declare a desired state (e.g., “3 replicas of nginx running”).
2. Kubernetes observes the actual state of the cluster.
3. A controller continuously compares desired vs. actual state.
4. If they differ, the controller takes action to reconcile them.
5. Repeat forever.

This “control loop” (or “reconciliation loop”) pattern is used everywhere in Kubernetes — it's not one feature, it's the architectural philosophy. Once this clicks, most of Kubernetes' behavior becomes predictable.

A Kubernetes cluster = a set of machines (nodes), split into two roles:

Control Plane (the brain) — makes global decisions, doesn't run your app containers:

• API Server: the front door. Every interaction (kubectl, controllers, dashboards) goes through it. It's the only component that talks to etcd.
• etcd: a distributed key-value store holding the entire cluster state. Think of it as the cluster's single source of truth/database.
• Scheduler: decides which node a new pod should run on, based on resource needs, constraints, and node availability.
• Controller Manager: runs the control loops (e.g., ensuring the right number of pod replicas exist).

Worker Nodes (the muscle) — where your actual application runs:

• Kubelet: an agent on each node that talks to the API server and ensures containers described for that node are actually running.
• Container Runtime: the software that actually runs containers (e.g., containerd).
• Kube-proxy: handles networking rules so traffic reaches the right pods.

Everything in Kubernetes is an object described in YAML/JSON, stored in etcd. Learn these in this order, smallest to largest:

Kubernetes objects don't reference each other by hard IDs — they use labels (key-value tags) and selectors (queries over labels). A Service doesn't say “send traffic to pod-123”; it says “send traffic to any pod labeled app: frontend.” This loose coupling is why pods can be destroyed and recreated constantly without breaking anything pointing at them.

Kubernetes assumes a flat network model with these rules:

• Every pod gets its own IP.
• Pods can talk to all other pods without NAT, regardless of node.
• Nodes can talk to all pods.

This is called the Kubernetes networking model, and it's implementation-agnostic — a plugin (CNI: Container Network Interface) fulfills these guarantees underneath. You don't need to master CNI internals early, but knowing this model exists explains why Services and DNS “just work” across nodes.

Theory sticks when paired with doing. Suggested order:

• Comfort with Linux command line
• Docker fundamentals: images, containers, Dockerfile, docker run, docker build
• Basic YAML syntax
• Basic networking: IP addresses, DNS, ports

Use a lightweight local cluster (don't start with cloud — remove billing/complexity while learning):

• minikube or kind (Kubernetes-in-Docker) — both run a real cluster on your laptop.
• Install kubectl, the command-line tool for talking to the API server.

1. Run a single Pod imperatively (kubectl run), inspect it (kubectl describe, kubectl logs).
2. Write a Pod YAML manifest by hand, kubectl apply -f, and delete/recreate it to see the lifecycle.
3. Create a Deployment; scale it up/down; kill a pod manually and watch it get recreated (this is where the control-loop theory becomes visible).
4. Expose the Deployment with a Service (ClusterIP, then NodePort) and understand why the Pod IPs alone weren't enough.
5. Add a ConfigMap and Secret; mount them into a pod as environment variables and as files.
6. Perform a rolling update (change the image tag) and a rollback (kubectl rollout undo).
7. Add a PVC to a pod and understand what happens to data when the pod is deleted vs. the PVC.
8. Set up an Ingress controller (e.g., ingress-nginx) and route traffic by hostname/path.
9. Explore resource requests/limits and watch what happens when you exceed them.
10. Look at a StatefulSet example (e.g., a small database) to see stable naming/storage in action.

• kubectl get, describe, logs, exec, port-forward, top — the daily-driver commands.
• Reading events (kubectl get events) to debug scheduling/crash issues.
• Understanding pod lifecycle phases: Pending → Running → Succeeded/Failed, and probes (liveness/readiness/startup).

• Helm: templating/packaging Kubernetes manifests (the “package manager” for K8s).
• Horizontal Pod Autoscaler: scaling based on metrics — another control loop.
• RBAC: who can do what in the cluster.
• Observability: Prometheus + Grafana for metrics, centralized logging.
• GitOps: tools like ArgoCD/Flux that apply the control-loop idea to deployment itself — git as the source of desired state.
• A managed cloud cluster (EKS/GKE/AKS) once local concepts are solid, to see how cloud integration (load balancers, storage classes) plugs into the same object model.

Every time you learn a new Kubernetes object, ask three questions:

1. What desired state am I declaring?
2. What actual state is being observed and compared against it?
3. What controller is reconciling the difference?

If you can answer those for Deployments, Services, PVCs, and HPAs, you understand Kubernetes' actual design — not just its API surface. Almost every “advanced” feature (operators, custom resources, autoscalers) is the same control-loop pattern applied to a new kind of desired state.

Keep coming back to this table as you go deeper — it's easy to get lost in YAML details and lose sight of the underlying “why.”