An experienced systems administrator who avoided Kubernetes to retain tight control of workload placement describes a hands-on conversion from skeptic to informed evaluator. The write-up defines core Kubernetes concepts (Pod, Node, Cluster, Service, Namespace) but emphasizes two structural strengths: control loops (Controllers) and Services. Control loops continuously drive system state toward declared goals using sensors and actuators, illustrated with queue-depth scaling, latency-driven horizontal scaling, and node health remediation. Services provide a stable virtual IP and DNS plus distributed load‑balancing via kube-proxy, which programs netfilter rules on every node so no single node is a forwarding bottleneck. That design yields simple implementation, cluster-scaled forwarding capacity and client robustness, but also lots of forwarding rules, potential double-forwarding during pod moves, and dependency on rule-evaluation performance.
Operational behavior of workloads and storage receives focused attention: Pods scheduled manually are fragile, ReplicaSets enforce replica counts, and Deployments enable eventual, controlled rolling updates by creating new ReplicaSets. Storage options contrast ephemeral volumes (pod-scoped) with PersistentVolumes claimed by PVCs; StatefulSets add per-replica stable identities and ordered lifecycle semantics that the author finds odd. The conclusion: the controller pattern and Service networking are sound architectural choices, but open questions remain about continuous instability of cluster state, richer autoscaling sensors, the out-of-tree CNI/CSI model versus cloud integrations, and practical PV inspection and operational workflows.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.