Deep Network GmbH Developers' Blog

Some notes on Kubernetes Networking

This is basically a summary of “Kubernetes Networking” chapter in “Container Networking: From Docker to Kubernetes”, as well as some relevant notes, investigations, and articles. The book is available here for free.

Kubernetes requires the following from the networking setup, the rest is up to you:

  • Containers can communicate with all other containers without NAT.
  • Nodes can communicate with all containers (and vice versa) without NAT.
  • The IP a container sees itself is the same IP as others see it.

There are 3 types of communication that can happen on the network layer: 1. container to container within a pod 2. pod to pod 3. to/from the cluster

#1 The infrastructure container or rather pause container is responsible of namespace sharing, including network, between containers within a pod. Therefore, all containers are on the same host and share the same IP. There is a good article here that goes into details of this special container.

#2 Every pod gets its own IP (that containers within that pod shares among each other as explained above) on the same network. A pod can talk to another pod directly without worrying about NAT. A detailed article on this is available here It is important to realise that your pods sit a different network than your nodes which is built on the node network.

Your service objects on the hand get a virtual IP which can be used to avoid dealing with pod IPs that might be changing as the pods get destroyed and created. These virtual IPs however exist only on the Kubernetes realm and are not real IPs on the network stack.

Let’s examine a real case:

  • We have an nginx statefulset with 2 instances.
    • pod 1 IP:
    • pod 2 IP:
  • We have an nginx service that maps these two instances
    • service IP:
  • We have two nodes, each hosting one of the pods

In one of the nodes, if we examine the iptables rules this is what we see the following lines:

  1. jump to KUBE-SVC-YK2SNH4V42VSDWIJ for anything that targets the service IP.
    -A KUBE-SERVICES -d -p tcp -m comment --comment "default/nginx:web cluster IP" -m tcp --dport 80 -j KUBE-SVC-YK2SNH4V42VSDWIJ
  2. Load balance the requests. Send half of them to KUBE-SEP-B4KBD76YHXMJL4VO and other half to KUBE-SEP-V73B5NNFXA6V3U7A.
    -A KUBE-SVC-YK2SNH4V42VSDWIJ -m statistic --mode random --probability 0.50000000000 -j KUBE-SEP-B4KBD76YHXMJL4VO
  3. Send KUBE-SEP-B4KBD76YHXMJL4VO to IP of Pod 1, and send KUBE-SEP-V73B5NNFXA6V3U7A to IP of Pod 2.
    -A KUBE-SEP-B4KBD76YHXMJL4VO -p tcp -m tcp -j DNAT --to-destination
    -A KUBE-SEP-V73B5NNFXA6V3U7A -p tcp -m tcp -j DNAT --to-destination


Jack of All Trades - Master of None