From One Crate to a Fleet (why Kubernetes)

The moment a single crate on one cart becomes a fleet too big for anyone to walk the docks and check by hand, and someone has to stand watch instead.

Before you start. Compose’s whole world was one docker-compose.yml, one command, one machine: docker compose up starts every container you listed and then it’s finished, done, out of the picture. Everything below asks what happens once that world is too many machines, and too many failures, for one command or one person to run by hand.

The idea in 30 seconds

Kubernetes runs a control loop, not a command: observe what’s running, compare it to what you declared, act to close the gap, forever.
The habit came from Google’s Borg (started around 2003), built because no team could restart failures across tens of thousands of machines by hand.
Three Google engineers open-sourced the idea in June 2014 and named it Kubernetes, Greek for “the helmsman.”
Self-healing, scaling and rollouts are the same loop, closing a different kind of gap each time.
The control plane - api-server, scheduler, controllers, etcd - is a handful of ordinary programs, not magic, and you can list all seven of them on a laptop cluster.
Kubernetes reconciles state, not events, so a lost message costs a delay, never a wrong answer. That’s called level-triggered.

Read the story · Skip to the numbers