Under the Hood (Kubernetes Internals)

Opens the control room drawn from the outside, and finally walks in. By the end, a stuck pod names the one program that stuck it.

Before you start. A pod is one running container, or a few that share a machine, that the cluster promises to keep running; The Harbor Master (pods, deployments, services) showed that, along with the other things you ask the cluster for by name: a Deployment (the app’s recipe), a ReplicaSet (the copy count) and a Service (one shared address). A small loop keeps comparing what you asked for with what is actually running and fixes the gap; Self-Healing & Autoscaling showed it at work. That loop is called the reconcile loop. This lesson opens the five ordinary programs that write those things down and run that loop. If copies of a store agreeing by vote is new, Who holds the pen (consensus & Raft) in System Design shows it in one read.

The idea in 30 seconds

Kubernetes is five ordinary programs: the API server (the one front desk), etcd (the one filing cabinet), the scheduler (picks the room), the controller manager (the supervisors who keep comparing plan to reality), and the kubelet (the caretaker on every machine).
Only the API server ever touches etcd. Everyone else watches the API server and acts on what it sees.
The scheduler decides where a job runs. It never starts anything itself.
The controllers run the reconcile loop: compare what a folder asks for against what’s true, and close the gap. That is what “self-healing” means in practice.
The kubelet is the one program that turns a decision written in a folder into a real Linux process on a real machine.
A stuck pod points at exactly one of these five, once you know their names.

Read the story · Skip to the numbers