I worked on two top-50 supercomputer clusters for science applications. You learn a lot by doing. You find out what happens to your electrical power, TFTP, DHCP, and syslog servers, and other infrastructure when those hundreds/thousands of nodes boot up all at once. Your Cisco switch reloads when you add a secondary IP address to all of your network interfaces because it's CAM table overflows. You learn how to automate without stuff like Ansible and Chef because they don't exist yet.
That's not to say this isn't cool engineering for engineering's sake. Just that if you are considering practical applications for this, maybe think twice.
Really? What software in this space (headless server cluster) is still x86 only?
I haven't checked recently, but at least that's what I remember.
The real selling point for me is energy efficiency and price, but the reason I could never make it happen is because most of the applications I want to use are x86 only and I don't have the resources to be a trailblazer in that area. In addition sourcing becomes an issue with raspberry pis when you aren't a big company like Oracle. It's hard to buy more than a handful as an individual.