Glad to hear you're thinking about this. Understanding the complexity provides power, but the tradeoff in this instance doesn't feel quite right.

For example, etcd is a powerful primitive, and then more complex/sophisticated systems can be built on top of it.

I wonder if an 80/20 solution that is simpler than fleet/systemd for pushing work into a CoreOS cluster would be a win, and then more complex systems (e.g. Kubernetes-esque orchestraction) could live on top of that.