In the real-world scenario, we’d probably have more than two worker nodes. Even a slight increase to three nodes would give us a very good chance that the failed server would be recreated in the same zone. Auto-scaling groups are trying to distribute EC2 instances more or less equally across all the zones. However, that is not guaranteed to happen. A good minimum number of worker nodes would be six.

The more servers we have, the higher are the chances that the cluster is fault tolerant. That is especially true if we are hosting stateful applications. As it goes, we almost certainly have those. There’s hardly any system that does not have a state in some form or another.

If it’s better to have more servers than less, we might be in a complicated position if our system is small and needs, let’s say, less than six servers. In such cases, we’d recommend running smaller VMs. If, for example, you planned to use three t2.xlarge EC2 instances for worker nodes, you might reconsider that and switch to six t2.large servers. Sure, more nodes mean more resource overhead spent on operating systems, Kubernetes system Pods, and few other things. However, we believe that is compensated with bigger stability of your cluster.