node name inconsistencies across platforms.

[hexfusion]

research

I ran into this in the past when I was trying to conclude a standard for etcd names. What I found is that some instances where using what is the equivalent of $(hostname), while others were using a concat of $(hostname) master-1 and ($hostname -d) ip-10-0-132-53.ec2.internal

Within a cloud instance, various status APIs exists to easily obtain metadata for the compute.

Example AWS.

$ curl http://169.254.169.254/latest/meta-data/hostname
ip-10-0-139-133.ec2.internal

So it appears that the kubelet is smart enough in some clouds to get what CEO expects which is a routable internal DNS name for the node name.

https://github.com/kubernetes/kubernetes/blob/38acec9bbc955a33c3366dc6082df90d18229b6f/staging/src/k8s.io/legacy-cloud-providers/aws/aws.go#L902

https://github.com/kubernetes/kubernetes/blob/38acec9bbc955a33c3366dc6082df90d18229b6f/pkg/kubelet/kubelet.go#L457

But then for metal, we just see master-1, yet it does have a routable FQDN.

oc get nodes                                               
NAME       STATUS   ROLES    AGE   VERSION
master-0   Ready    master   15h   v1.17.1
master-1   Ready    master   15h   v1.17.1
master-2   Ready    master   15h   v1.17.1
worker-0   Ready    worker   14h   v1.17.1

$ dig +short master-0.ostest.test.metalkube.org
192.168.111.20

proposed resolution

Non clouds use --hostname-override which is populated by ENV.

ExecStart=/usr/bin/hyperkube \
    kubelet \
      --config=/etc/kubernetes/kubelet.conf \
      --bootstrap-kubeconfig=/etc/kubernetes/kubeconfig \
      --kubeconfig=/var/lib/kubelet/kubeconfig \
      --container-runtime=remote \
      --container-runtime-endpoint=/var/run/crio/crio.sock \
      --node-labels=node-role.kubernetes.io/master,node.openshift.io/os_id=${ID} \
      --minimum-container-ttl-duration=6m0s \
      --hostname-override=${KUBLET_HOSTNAME_OVERRIDE} \
      --cloud-provider=aws \
      --volume-plugin-dir=/etc/kubernetes/kubelet-plugins/volume/exec \
       \
      --register-with-taints=node-role.kubernetes.io/master=:NoSchedule \
      --v=${KUBELET_LOG_LEVEL}

I understand we can change our code to reflect master-1 as ETCD_NAME. But I feel standardizing node naming standard might be a good move across the product.

[deads2k]

I think we can move away from dns names entirely. By relying solely on internal IPs, we can have the operator avoid the dns dependency and conflicting values like this entirely.