Limits of a kind Cluster

The runs above prove that the submission path, pod lifecycle and dynamic allocation work, and nothing more. A single kind node on a 4-CPU workstation that two other workloads share cannot show what makes Kubernetes 5,150 worth its complexity for Spark 129 : many nodes, an autoscaler, executors spread across failure zones, object storage instead of a host mount, and a shuffle that crosses a real network. Its timings would only mislead, and the operators, pod templates, schedulers, and the security set-up (namespaces, network policies, secrets) would each need a production-like cluster to demonstrate honestly. So the chapter treats Spark on Kubernetes as configuration you can read and verify, with one small real run, and leaves operating it to Package Managers and DevOps's Kubernetes chapters and to the managed platforms of Managed Spark Platforms, several of which run Spark on Kubernetes for you.