Inside the infrastructure layer · novice → expert

The cluster that
heals itself.

diagram first — always

You don't tell Kubernetes what to do — you tell it what you want, and it works, forever, to make reality match. A node dies, a pod crashes, traffic spikes: the cluster notices and repairs itself. This is that machine, drawn before it's explained.

See it reconcile Start with infrastructure
Anatomy

Two halves: a brain that decides, and muscle that runs.

Every cluster splits in two. The control plane holds the desired state and makes decisions; the data plane — your worker nodes — actually runs the containers. Everything Kubernetes does is a conversation between these two.

Control plane

The brain

Decides what should be true and never stops checking.

storeetcdThe cluster's single source of truth — every object, the whole desired state.
gatewayAPI serverThe one front door. Everything reads and writes cluster state through it.
placementschedulerPicks which node each new pod should run on, by fit and constraints.
reconcilecontrollersWatch for drift and act to close the gap between desired and actual.
Data plane

The muscle

Where your containers actually live and breathe.

agentkubeletOn every node — starts pods the API server assigns and reports health back.
runtimecontainer runtimecontainerd / CRI-O — the thing that actually pulls images and runs containers.
networkkube-proxyWires up service networking so a stable IP reaches the right pods.
unitthe podOne or more containers that live and die together — Kubernetes' atom of work.
The flow

From kubectl apply to running pods.

One command sets a chain in motion. Follow it across the cluster — this loop is the whole idea of Kubernetes, running thousands of times a second.

01 · declare

You state the want

“Run 3 replicas of this image.” It lands in etcd via the API server as desired state.

02 · notice

A controller sees a gap

Desired says 3, actual says 0. The deployment controller creates the missing pod objects.

03 · place

The scheduler assigns

Each unscheduled pod is matched to a node with room and the right constraints.

04 · run

The kubelet obeys

On each node, the kubelet pulls the image and tells the runtime to start the container.

05 · watch

It never stops

A pod dies? Actual drops to 2, the gap reopens, and the loop heals it. Forever.

Why it's different

Once you see the loop, Kubernetes stops being a maze of YAML.

01 / see

One idea, drawn

Reconciliation is the whole engine. See it once as a diagram and every controller, operator and CRD becomes the same shape.

02 / connect

Sits inside infrastructure

This is a deep dive under the infrastructure layer of the stack — containers scheduled, routed and scaled, made legible.

03 / deepen

Debug with a model

When something's stuck, you'll know which half to look at — is desired state wrong, or is the data plane failing to reach it?

Begin

Watch a cluster heal itself.

Declare a want, break something, and see the reconcile loop put it right. That's Kubernetes — no magic, just a loop that never quits.

See it reconcile