I’m still a noob with Kubernetes architecture, but I’m beginning to have a strong opinion that it is better to think early about what the clusters (plural) will eventually be in your environment and not the cluster (singular). The place I work now has really worked themselves into a corner by not anticipating the need for multiple clusters later. Small decisions, like maintaining an external etcd, are now impacting the ease with which we deploy additional clusters for other reasons.
Even though Kubernetes takes care of Node affinity and has taints and tolerations to help isolate Nodes during selection, the unnecessary complication of this when deploying to Nodes on a vastly heterogeneous hardware collection adds unnecessary overhead and risk. It seems better to think of the clusters in terms of hardware similarities and to build them that way from the very beginning.
I’m still unsure because the work of setting up and maintaining an HA control plane isn’t trivial and burns several machines in the process (which is a factor at home when I don’t have that many to begin with) but I’m still going to think about multi-cluster from the beginning more.