Thanks for that detailed explanation. The fact that all main functions live in a single process probably helps a lot with CPU and RAM usage, compared to k8s where etcd, API server, scheduler and controller manager live in independent processes, which means serialization/deserialization and less optimization opportunities.