From @errordeveloper on October 13, 2016 9:18

This makes sense to add. Are you envisioning re-joining another cluster or simplify clean-up and recycle nodes? Right now you can just throw away a node and run kubectl delete node <name> to deregister it.

From @pires on October 14, 2016 20:0

@luxas please assign this to me.

From @pires on October 14, 2016 20:55

I'll add this behavior after #34807 lands, such that both
kubeadm reset and kubeadm leave result in the node been marked as unschedulable to the cluster.

Thoughts @luxas @errordeveloper?

From @pires on November 8, 2016 9:22

This is a list of potential execution steps:

• Mark current node as unshedulable aka cordon the node
• Drain the node
• Restart the kubelet with empty or default values, e.g. crash-loop until an API server is provided

1 - Existing kubectl code living in pkg/kubectl/cmd/drain.go should be either re-used or adapted.
2 - kubeadm reset should execute similarly, differing only in the last step where instead of restarting, it should stop the kubelet.

Thoughts?

Restart the kubelet with empty or default values, e.g. crash-loop until an API server is provided

I raised discussion about how to change kubelet flags in #64.

Really need thoughts before proceeding.

We need to discuss what the semantics would be regarding leave vs reset.
Could we just fix reset to actually remove the node from the cluster as well (right now the node will be left in kubectl get no with status NotReady)

What would the use-case for leave be that reset wouldn't do?

@luxas leaving a cluster is different than resetting everything, including stopping the kubelet. But I agree both should perform node cordon and drain.

@pires so when would you leave but not reset?
That's what I want to know...

Are you thinking about something like a node restart? But that can be fixed with a plain kubectl command...
I just want to figure out the use case for leave

@luxas i joined master 1. now i want my node to join a different master. Leave is just reset but with keeping the system ready to re-run kubeadm join. More importantly to me, reset needs to cordon and drain.

To me it sounds like you want:

• reset to do cordon and drain (which I'm a huge +1 on, I just haven't had that extra time)
• leave to just run reset and systemctl start kubelet

And it sounds like leave then is unnecessary (unless we decide to rename reset to leave indeed)
We could consider reset to run systemctl start kubelet as well, but I think we should prefer to have as little interaction with systemd as possible. In fact, I would be more in favor for automatically trying to start the kubelet in that case.

I agree that 'leave' doesn't seem necessary. I would rather contemplate
'join --reset'.

On Wed, 23 Nov 2016, 17:43 Lucas Käldström, notifications@github.com
wrote:

To me it sounds like you want:

• reset to do cordon and drain (which I'm a huge +1 on, I just haven't
  had that extra time)
• leave to just run reset and systemctl start kubelet

And it sounds like leave then is unnecessary (unless we decide to rename
reset to leave indeed)
We could consider reset to run systemctl start kubelet as well, but I
think we should prefer to have as little interaction with systemd as
possible. In fact, I would be more in favor for automatically trying to
start the kubelet in that case.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#42 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AAPWS727rwkOLFI2pxH-z8bEu-AKJBQdks5rBHs2gaJpZM4K5xyN
.

@errordeveloper I still want the top-level command, but I think we can just do fine with reset (or leave if we want to call it that)

Feel free to reassign this to anyone else.

reset is applicable for both master and slave nodes, where leave mostly for non-master nodes (where graceful shutdown and removing from cluster is needed)

The PR is here: kubernetes/kubernetes#37831

Fixed with kubernetes/kubernetes#37568

@luxas kubeadm reset does not delete the network interfaces. It does delete the /etc/cni/net.d though. Should it not do a complete reset? I did raise an issue with weavenet here, but it is more of a kubeadm cleanup issue.

@cheburakshu Cleaning up the network is hard; you're free to install whatever networking solution you want, so therefore kubeadm can't know how to cleanup something it hasn't installed.

Unfortunately this is out of scope for kubeadm, we can't build in cleanup solutions for every single network provider.