DO node metadata InternalIP is set incorrectly

[dghubble]

Node metadata InternalIP is correct on AWS, Azure, GCP, and bare-metal, but regressed on DO in v1.12.3.

Historically, DigitalOcean node registration has been the troublesome one. Originally, DigitalOcean kubelets used --hostname-override to register with an internal IP to workaround DO not resolving hostnames like other clouds. Later, a feature request asked for friendly node names (e.g. worker-1, not IP). A (still legitimate) change to apiserver addressing preference was made across platforms. That (seemingly) allowed DO's override to be dropped and pretty names used. This did cause a regression in etcd scraping on DO, which was fixed before v1.12.3.

But the original behavior was correct. When DO registers its hostname (e.g. nemo-controller-0), unlike other cloud providers, DO does not resolve that to an internal IP. Instead, Kubelet has to fallback to using the IP of the default gateway, which is the public IP. Preferred addressing is irrelevant.

I'll fix this weekened. Overall, its not impacting to cluster usability, firewall rules are by tag (source) which is why its fine. But DO should go back to registering kubelets via internal IP for correctness.