DO/bare-metal: Change preferred-address-type instead of overriding hostname

[klausenbusk]

Feature Request

On DO/bare-metal we currently override the hostname with the private ip (relevant commit), as the hostname isn't resolvable:

• Kubelets should register nodes via their private IPv4 address,
  as provided by the metadata service from Digital Ocean
• By default, Kubelet exec's hostname to determine the name it should
  use when registering with the apiserver. On Digital Ocean, the hostname
  is not routeable by other instances. Digital Ocean does not run an
  internal DNS service.
• Fixes issue where the apiserver can't reach the worker nodes. This
  prevented kubectl logs and exec commands from working

Feature

It would be more optimal to use the real hostname as it is easier to identify a node by its hostname, rather than its ip.
We can archive that by changing (kube-apiserver):

  --kubelet-preferred-address-types strings                 List of the preferred NodeAddressTypes to use for kubelet connections. (default [Hostname,InternalDNS,InternalIP,ExternalDNS,ExternalIP])

Which change how the apiserver connect to the nodes. I'm currently using --kubelet-preferred-address-types=InternalIP,Hostname,ExternalIP on a regular bootkube DO cluster without any issues.

Tradeoffs

I'm not sure how this will affect AWS/GCE clusters.

[justaugustus]

I ran into some woes with this flag in Tectonic, and ended up going with --kubelet-preferred-address-types=InternalIP,ExternalIP,Hostname.
I can confirm that works (in Tectonic, at least) across AWS, Azure, and BM.

Check out the convo here: coreos/tectonic-installer#2777.

[dghubble]

Note, this broke the Prometheus etcd scrape on DigitalOcean clusters. Fixed by 7de03a1