kubernetes: endpoints for daemonset in host network not ready and inconsistent with pod IPs

Is this a BUG REPORT or FEATURE REQUEST?: /kind bug

What happened: After deploying the prometheus node-exporter as daemonset + service with these manifests: https://gist.github.com/discordianfish/77b1e4f82ebe966ec5ce341aa3366155

…the node-exporter gets deployed to all nodes as expected:

$ kubectl get pod -l app=node-exporter -o json|jq '.items[]|{"name": .metadata.name, "ip": .status.podIP}'|jq -s 'sort_by(.ip)'
[
  {
    "name": "node-exporter-2b14f",
    "ip": "10.32.130.2"
  },
  {
    "name": "node-exporter-nkqhr",
    "ip": "10.32.130.3"
  },
  {
    "name": "node-exporter-7w9pq",
    "ip": "10.32.130.4"
  },
  {
    "name": "node-exporter-j3hfm",
    "ip": "10.32.130.5"
  },
  {
    "name": "node-exporter-pp1sq",
    "ip": "10.32.130.6"
  }
]

The pod IPs match the host ips, which I assume is expected given the node-exporter runs in host networking namespace:

$ kubectl get nodes -o json| jq '.items[].status.addresses[]|select(.type == "InternalIP")'
{
  "address": "10.32.130.6",
  "type": "InternalIP"
}
{
  "address": "10.32.130.5",
  "type": "InternalIP"
}
{
  "address": "10.32.130.4",
  "type": "InternalIP"
}
{
  "address": "10.32.130.2",
  "type": "InternalIP"
}
{
  "address": "10.32.130.3",
  "type": "InternalIP"
}

So far so good. The problem is that all but one endpoints are ‘not ready’ and they are all inconsistent with their pod IP:

$ kubectl get endpoints -l app=node-exporter -o json | jq '.items[0].subsets[0].addresses[]|{"name": .targetRef.name, "ip": .ip}'
{
  "name": "node-exporter-pp1sq",
  "ip": "10.32.130.6"
}
$ kubectl get endpoints -l app=node-exporter -o json | jq '.items[0].subsets[0].notReadyAddresses[]|{"name": .targetRef.name, "ip": .ip}'
{
  "name": "node-exporter-2b14f",
  "ip": "10.32.130.3"
}
{
  "name": "node-exporter-nkqhr",
  "ip": "10.32.130.4"
}
{
  "name": "node-exporter-j3hfm",
  "ip": "10.32.130.5"
}
{
  "name": "node-exporter-7w9pq",
  "ip": "10.32.130.7"
}

This leads prometheus to miss one instance and apply wrong labels. This could lead to misinterpretation of monitoring results in this particular case and probably even catastrophic failure in other cases.

What you expected to happen: I expected the endpoint IP match the pod_ip.

How to reproduce it (as minimally and precisely as possible): This is a pretty much vanilla GKE cluster with https://gist.github.com/discordianfish/77b1e4f82ebe966ec5ce341aa3366155 applied.

Anything else we need to know?: The node-exporter is running in host networking and pid namespace.

Environment:

• Kubernetes version (use kubectl version): 1.6.4
• Cloud provider or hardware configuration**: GKE