Let's say you run kubectl get pods and something, somewhere, answers. Here's who.

A cluster is a control plane plus one or more worker nodes, say the docs. The control plane holds the API server everything talks to, etcd for state, the scheduler, the controllers. Each node runs the kubelet keeping pods alive, kube-proxy, a runtime. Thinking kubectl talks to your containers is the first leftover; it talks to the API.

Machines below, control plane above. You talk to the API.

Let's say kubectl apply -f feels like uploading a config file. It's stranger and better.

An object is a record of intent: once created, the system constantly works to ensure it exists, say the docs. Your YAML becomes an object in etcd, a standing order the cluster keeps chasing. Thinking of apply as a script that ran and finished is the model that makes the rest confusing.

You declare. The cluster ensures. That's the trick.

Let's say a pod dies and a new one appears, and nobody ran a command. Somebody was chasing.

The spec is the desired state you set; the status is what the system reports back, say the docs. Ask for replicas: 3 and the system starts three, records what it finds, replaces any that die. Editing the status to fix a problem is painting the smoke detector. The spec is yours.

You declare. The controller does the chasing.

Let's say you deploy one container and kubectl answers in pods. That box is the unit.

A Pod is the smallest deployable unit: containers sharing storage and network, co-located, co-scheduled, say the docs. One IP for the whole pod, containers talking over localhost. One container per pod is the common case. Squeezing loosely-related containers into one pod couples their lives for no prize.

Kubernetes doesn't run your container. It runs a pod around it.

Let's say your app calls a pod IP that worked yesterday. Pods don't keep addresses. Something has to.

A Service exposes a group of pods behind one stable endpoint, updated whenever the pods change, say the docs. The IP and DNS name stay put while the pods behind them rotate. Hardcoding pod IPs in config is the bug this object exists to delete.

A Service is an address, not a box.

Let's say you binge this book in one sitting and by Thursday it's fog. That's the forgetting curve, not a character flaw.

Spaced practice beats cramming: Cepeda et al. 2006, a meta-analysis of 317 experiments (PubMed 16719566), the one practice claim this book allows. So the convention, named as a convention: one object a day, one command, on a real cluster. The Saturday binge feels productive and doesn't survive the week.

A little, often, on a real cluster. That's the method.

Let's say it clicked somewhere around the Service page and you want the whole machine in one look.

One declaration: Deployment for copies, labels wiring the Service, ConfigMap and Secret injecting, Ingress at the door, probes for health, requests and limits for the node, HPA for the count, every piece from the docs. The reconcile loop under it all holds it true. Learning commands before this shape is sentences in a language you can't yet speak.

Declare the objects. The loop holds them true.