Fixes issue with etcd DNS resolution via locally provided nameserver.

[shift]

Issue

When restarting masters, etcd-member.service fails to be able to reverse lookup the names of the TLS nodes as DNS hasn't been defined yet. As a side effect causes some issues with CLUO deployed ;)

Dec 04 07:27:16 node0.cluster.com etcd-wrapper[875]: 2017-12-04 07:27:16.092239 I | etcdmain: rejected connection from "192.168.15.13:44308" (tls: "192.168.15.13" does not match any of DNSNames ["node0.int.cluster.com" "node3.int.cluster.com" "*.kube-etcd.kube-system.svc.cluster.local" "kube-etcd-client.kube-system.svc.cluster.local"])

I've only used this on bare-metal so not sure if it effects other module types.

Change Details

This change makes it so that the wait-for-dns.service uses the RequiredBy directive for etcd-member.service.

Testing

I've used this fix previously from prior to Typhoon and has been running successfully in more than one cluster.

[dghubble]

Thanks! I've seen this race occur on rare occasions, but didn't get to changing it. If etcd-member.service starts before /etc/resolv.conf is populated the container is left with an empty /etc/resolv.conf and cannot resolve its peers. The quick fix is to systemctl restart etcd-member.servce.

Its unfortunate, but I think having both kubelet.service and now etcd-member.service require the little wait-for-dns.service shim unit is a suitable fix. It can occur on any platform so the fix should be applied in all controller container linux configs. Want to tweak the others?

[shift]

Updated PR to include all 4 profiles :)

[dghubble]

Thanks!

[dghubble]

Merged as ce49a93 to normalize commit message