Consider using kubeadm for Kubernetes cluster setup

[resouer]

The benefit of kubeadm is it will be responsible for handling Kubernetes versions. i.e. versions of etcd, kubelet, kube-addons, master pods, also it support custom CRI, which is required by GPU.

The change to update to kubeadm is straigtforward:

1. Install kubeadm and kubelet on every node.
2. On master node, instead of create master Pods, just execute kubeadm init
3. On every worker node, execute kubeadm join <masterIP:port>
4. Intall network plugin: kubectl apply -f flannel.yaml
5. Install addons: dns is included by default, others are the same as today

No enforcement here, just brought up a thought.

[resouer]

cc @luxas, correct me if anything wrong.

This is the AI infrastructure of Microsoft Research, which is a perfect candidate of using kubeadm I think.