Kubernetes

The de-facto container orchestration platform and the substrate of most modern platform-engineering. In this wiki it appears not as a tutorial subject but as the operational reality teams run in production — the assembly point for the wider CNCF ecosystem (Prometheus, Cilium, cert-manager, Cluster API, GitOps).

What it is

Per the first-party docs, Kubernetes is “a portable, extensible, open source platform for managing containerized workloads and services, that facilitates both declarative configuration and automation.” Google open-sourced it in 2014, distilling 15+ years of running production workloads on its Borg system (the name is Greek for helmsman; K8s = the 8 letters between K and s). The load-bearing idea for this spoke is the continuous control loop: you declare desired state and Kubernetes works to maintain it — self-healing (restart/replace/reschedule failed containers), automated rollouts/rollbacks, bin packing onto nodes, horizontal scaling, service discovery + load balancing, and secret/config management.

Just as telling is what it is not: not an all-inclusive PaaS. It doesn’t build or deploy your source, ships no built-in database/middleware, and mandates no logging/monitoring/alerting — which is precisely why production Kubernetes is an integration problem (kubernetes-integration-tax): the platform leaves the observability, networking, and delivery layers for you to wire together.

In the sources

kubernetes-integration-tax is the anchor: production Kubernetes is dominated by the hidden labour of integrating its surrounding tools, not by Kubernetes itself. This is where the observability stack (Prometheus + Cilium metrics) and networking (NetworkPolicies, ingress) actually have to be wired together.

The scheduler as a scaling limit

The first source to argue against Kubernetes on its own ground is modal-1m-sandboxes (T3, and self-interested — Modal sells the alternative). Its charge is specific: the scheduler is O(n × p) over nodes × pods and serialized by default, etcd turns into a bottleneck as pod writes accumulate, and node heartbeats levy a baseline O(nodes) write load before any workload exists. The control loop that makes Kubernetes self-healing — declared state, reconciled against a consistent view — is exactly what a burst of a million second-lived sandboxes can’t afford to pay for. This isn’t a defect so much as a workload boundary: the consistent cluster view earns its cost for long-lived services with real placement constraints, and stops earning it when the unit is ephemeral and the arrival is a spike (container-scheduling).

Adjacency

Kubernetes-as-a-rented-service (managed K8s pricing) would belong in cloud-wiki, but the platform-ops practice of running it in production lives here.

Related

kubernetes-docs-overview · webernetes · platform-engineering · platform-ops · kubernetes-integration-tax · observability · gitops · opentelemetry · prometheus · container-scheduling · modal-1m-sandboxes