Kubernetes and Containers

The single most important idea in Kubernetes is the reconciliation loop. You declare a Desired State Configuration, the cluster records it, and a controller continuously compares desired state against actual state and acts to close the gap. The Controllers and Reconciliation Loops lessons explain this engine in detail. If you ask for three replicas and one crashes, the controller notices two are running and starts a third. Nothing intervenes manually. The loop just runs forever.

Different workload shapes get different controllers. DaemonSets put exactly one copy on every node, which is how log collectors and node agents work. StatefulSets give each replica a stable identity and its own storage, which is what databases and queues need. Jobs and CronJobs run work to completion on a schedule rather than running forever. Init Containers run setup steps before the main container starts. Each is a variation on the same control loop with different guarantees.

When the built-in resources are not enough, you extend Kubernetes itself. Custom Resource Definitions let you teach the cluster a new object type, and Operators pair a CRD with a controller that knows how to manage a specific application, like running a Postgres cluster the way a human expert would. This is how the platform grows without forking it.

Google runs nearly everything on a container orchestrator and built Kubernetes from the lessons of its internal Borg system, which is why the open source project maps so closely to how hyperscale infrastructure actually works. The control loop, the scheduler, and the desired-state model all come straight from running planet-scale services.

Spotify, Airbnb, and most large engineering organizations use Kubernetes to give every team a self-service platform. A team ships an image and a declaration, and the cluster handles placement, scaling, restarts, and rollouts. The Operator pattern shows up everywhere too: companies package their stateful systems, from databases to message brokers, as Operators so on-call engineers are not manually running recovery steps at 3 AM.

The common thread is use of capacity, not micromanagement of machines. Whether you run ten services or ten thousand, the model is identical: declare what you want, let controllers reconcile it, and treat infrastructure as immutable and replaceable. That uniformity is why the same skills transfer across almost every company running modern backend systems.