Contrary to what is possible in a Docker Swarm, you cannot run containers directly in a Kubernetes cluster. In a Kubernetes cluster, you can only run pods. Pods are the atomic unit of deployment in Kubernetes. A pod is an abstraction of one or many co-located containers that share the same Kernel namespaces, such as the network namespace. No equivalent exists in the Docker SwarmKit. The fact that more than one container can be co-located and sharing the same network namespace is a very powerful concept. The following diagram illustrates two pods:
In the preceding diagram, we have two pods, Pod 1 and Pod 2. The first pod contains two containers, while the second one only contains a single container. Each pod gets an IP address assigned by Kubernetes that is unique in the whole Kubernetes cluster. In our case, these are the IP addresses 10.0.12.3 and 10.0.12.5. Both are part of a private subnet managed by the Kubernetes network driver.
A pod can contain one to many containers. All those containers share the same kernel namespaces, and in particular they share the network namespace. This is marked by the dashed rectangle surrounding the containers. Since all containers running in the same pod share the network namespace, each container needs to make sure to use their own port since duplicate ports are not allowed in a single network namespace. In this case, in Pod 1, the main container is using port 80 while the supporting container is using port 3000.
Requests from other pods or nodes can use the pod's IP address combined with the corresponding port number to access the individual containers. For example, you could access the application running in the main container of Pod 1 through 10.0.12.3:80.