One scenario is that when you rollout a new version, the old container will only be killed when new one is up and running. If there are many replicas running, Kubernetes will replace them with new versions one by one. And all of this behaviour is highly configurable.
[1] https://kubernetes.io/docs/tutorials/kubernetes-basics/updat...
It would be interesting to know how this works under the hood though, is this just built on top of docker swarm load balancing?
Kubernetes architecture is entirely different from that of Docker Swarm.
A Kubernetes Service object [1] can load balance traffic to various Kubernetes Pods (think containers) [2] as defined by it's " pod selectors". It can choose to direct traffic only to running containers.
[1] https://kubernetes.io/docs/concepts/services-networking/serv... [2] https://kubernetes.io/docs/concepts/workloads/pods/pod/