FWIW, I run Minikube to manage a single ML churning machine in my garage, because I only need to learn 'docker build Dockerfile', 'docker push <image>', 'docker run <image>', 'kubectl create -f <foo.json>' and 'kubectl delete -f <foo.json>' to manage the workloads on a regular basis. For this minimal brain-space investment I get a bunch of features like resource management, a workload queue, dashboards and being able to test/debug an image on my dev machine for free. Then I scale this knowledge up to managed clusters in the cloud with as little machines I can get away with, using minimal brain-space to learn additional systems.

While I agree with the general sentiment of using the bare minimum to get the job done, the gratuitous complexity problem is usually caused by the people using the tools than by the tools themselves.