Kubernetes is a desired-state orchestration system. You declare workloads and their constraints; controllers and schedulers try to make the cluster match that state.
Do not use “Kubernetes” as the explanation for every deployment behavior. Name the object or controller that owns the decision.
In this chapter
- Pods, nodes, and DeploymentsSeparate the machine, the smallest deployable unit, and the controller maintaining replicas.
- Services and networkingGive changing Pods a stable way to be reached.
- Storage and init containersPrepare shared data before the main application starts.
- Startup, readiness, liveness, and rolloutsKeep initialization, traffic eligibility, restart decisions, and version replacement separate.
- GPU scheduling, Pending Pods and inference incidentsRead scheduler symptoms, resource requests, model startup and GPU metrics.