Doing things like kubectl replace requires a lot of logic in kubectl, such as waiting for RC to scale pods down by polling before deleting the RC itself.

Putting this in kubectl makes it hard to replicate in clients written in other languages to get consistent behaviour. Moving this to the api server would mean consistency across languages.

For background, we created the Java Kubernetes client in fabric8 (https://github.com/fabric8io/kubernetes-client) & I've also contributed to the Ruby client at https://github.com/abonas/kubeclient). Having consistent behaviour means replicating undocumented functionality across clients which seems a bit dirty.