CCM needs agent/worker nodes to create the automatic loadbalancer for a service/ingress

[ikatergaris]

I suppose the easiest way to reproduce, would be to create a single node cluster with K3s and make it a master. Apply the generated ccm-linode.yaml following the howto and then try to LoadBalance a service. It will fail. Add an agent node to the cluster and then it will succeed.

If I'm correct about this, it is limiting our options.
For example, I have setup a HA K3s for rancher and planning to use it, to create more clusters. The 3 control plane nodes should be able to do the whole job. There should be no need to add agent nodes, because you would also need at least 2.

I thought I'd just reach out and see if this is something I'm doing wrong and there is a label, or an annotation I'm missing, that might solve my problem.

[ikatergaris]

This recently closed issue, contains the log message, when no agents have been added to the cluster. Hopefully it can help:
#68 (comment)

[sibucan]

Hi @ikatergaris! Thank you for bringing this up.

I think this was an issue in the upstream service controller, not in our code. It seems that, by design, a LoadBalancer service would not include nodes that are marked as being control plane (master) nodes, see: kubernetes/kubernetes#65618

For thoroughness, I checked the upstream service controller for our k8s version in go.mod (v1.19.2, yikes!) and yes -- it does exclude nodes if they are master nodes: https://github.com/kubernetes/kubernetes/blob/f5743093fd1c663cb0cbc89748f730662345d44d/staging/src/k8s.io/cloud-provider/controllers/service/controller.go#L641

This code seems to have been removed entirely starting in v1.21.0, so I believe the first thing we can do on our end is to update our k8s package to at least v1.21.0. From your end, you could try deleting the node-role.kubernetes.io/master label from your node to get the CCM to observe it for the LoadBalancer. Let me know if that workaround helps you for now while we get a new CCM release out the door.

[ikatergaris]

Your suggestion worked. I'll be eagerly waiting for the next release :-)
Thanks for the quick reply and workaround

[sibucan]

@ikatergaris The changes have been merged, let me know if you're able to add workers to your cluster w/o needing to delete that special label. 😄