I am running a Kubernetes cluster on GKE and I noticed that in kube-system the IP addresses of pods with
fluentd-gcp-...
kube-proxy-gke-gke-dev-cluster-default-pool-...
prometheus-to-...are the same as those of the nodes, while other pods such as
event-exporter-v0.3.0-...
stackdriver-metadata-agent-cluster-level-...
fluentd-gcp-scaler-...
heapster-gke-...
kube-dns-...
l7-default-backend-...
metrics-server-v0.3.3-...e.g.
kube-system fluentd-gcp-scaler-bfd6cf8dd-58m8j 1/1 Running 0 23h 10.36.1.6 dev-cluster-default-pool-c8a74531-96j4 <none> <none>
kube-system fluentd-gcp-v3.1.1-24n5s 2/2 Running 0 24h 10.10.1.5 dev-cluster-default-pool-c8a74531-96j4 <none> <none>where the pod IP range is: 10.36.0.0/14
and nodes are on 10.10.1.0/24
have IP addresses in pod address range. What is specific about the first three?
This is because pods such as kube proxy , Fluentd, Prometheus are running in host network directly via hostNetwork: true. You can describe those pods and verify that hostNetwork: true is present.
Now coming to the point why these pods need to run in host network in the first place , kube proxy needs access to host's IP tables, prometheus collects metrics and Fluentd collects logs from the host system.
You can just deploy a sample pod such as nginx with hostNetwork: true and it will get node IP.If you remove hostNetwork: true it will get IP from pod CIDR range.
apiVersion: v1
kind: Pod
metadata:
labels:
run: nginx
name: nginx
spec:
containers:
- image: nginx
name: nginx
restartPolicy: Always
hostNetwork: true