Master Assigned to Worker Nodes about kube-karp HOT 10 CLOSED

I had the same issue. Adding this to the end of values.yaml solves the issue for me.

affinity:
  nodeAffinity:
    requiredDuringSchedulingIgnoredDuringExecution:
      nodeSelectorTerms:
      - matchExpressions:
        - key: "node-role.kubernetes.io/master"
          operator: Exists

from kube-karp.

Yes, you can definitely do it with the Helm chart!

Edit the values.yaml here: https://github.com/immanuelfodor/kube-karp/blob/master/helm/values.yaml

nodeSelector:
    kubernetes.io/role: master

The tolerations in place allow the pods to run on master nodes, the node selection can restrict the pods only to the master nodes.

I only have clusters with master nodes, so I can't test this, please report back if it doesn't work.

Reference: https://stackoverflow.com/questions/60404630/kubernetes-daemonset-only-on-master-nodes

from kube-karp.

Thanks for responding swiftly!!

I know that's a possibility via nodeSelector but I was under the assumption that CARP must be active on all nodes, no? I should have clarified that in the original post, sorry.

from kube-karp.

That label doesn't seem to work anymore. I've used it once before but when I try installing the Helm chart with that label for nodeSelector, no pods are spawned.

Taking a look at the master labels, it seems it's been replaced with these:

It looks like their values are blank. I'm trying to get it functional

from kube-karp.

Yes, it seems my cluster also doesn't have that, this happens when blindly trusting online resources 😃

I have node-role.kubernetes.io/controlplane: "true" labels that I could use in my cluster.

Check what labels and values do you have with kubectl describe no NODENAME -o yaml. Check master and worker nodes as well, should you find a label that is only present on master nodes, you can use that. You can also label any node with anything then use that in the selector.

from kube-karp.

Thanks! I ended up just assigning an arbitrary label (color=red) and using that for nodeSelector. It seems this is somewhat of a common issue. I tried using the operator but that didn't work.

However, now my 3 master nodes are all assigned the vIP and in MASTER state...is that expected behavior?

from kube-karp.

Hmmm, that shouldn't happen, VIP management is not even related to k8s labels. Master nodes are all within the same subnet? Can master nodes communicate with the general multicast IP addresses to advertise themselves? See https://github.com/lorf/UCarp MULTICAST IP SELECTION section for more info. This issue seems to be related to network, the wrapped ucarp binary is probably not able to communicate to other pod's binaries, so they all think they are alone, so they all elect themselves as master and assign the VIP.

from kube-karp.

Yes I agree with you; likely a networking issue somehow. They are all in the same subnet. Thanks for linking and pointing that out. I will continue trying to solve this and post back.

from kube-karp.

Is tcpdump -n net 224.0.0.0/24 a surefire way to tell if the host can communicate with the multicast IPs?

If so, no packets captured. Even when not using nodeSelector.

from kube-karp.

Honestly, I never needed to debug multicast packets

from kube-karp.