If Cluster is used for redundancy, PXE server can be a weak link. I’m unsure on that for myself. But having no local boot drive is very very nice indeed. And as you said, super awesome tinkering stuff to learn and be proud on your achievement once everything is running 
But I’d gladly trade eMMC for virtually anything if I’m given a choice.
Welcome to L1T, grab a beer, we’re all guilty of too expensive toys, but we also deserve them
It’s nice to see things translating into practice. And also to see things that may be of more interest to you in the coming semesters.
Ceph, GlusterFS, or BeeGFS. All distributed storage/filesystems. Longhorn may be interesting too, if you are digging into Kubernetes at some point. Some may need some compiling and extensive tinkering, because ARM and not x86, but that translates also well into learning, even if it can be tedious and frustrating.
I know Jeff and 45 Drives had/have made Ceph work on Pis and Ceph has an ARM version (totally no clue how good or mature it is).
I picked up the term “Level-2 fun”. It’s the kind of fun that is often painful while you do it, but the reward at the end is worth it all and more. And getting code and stuff to work is sometimes like beating a Dark Souls Boss or physically exercising .
Always good to burn CPU cycles for science. I’m pretty sure you’ll find a nice project to keep your CPUs busy. Not sure on the GPU/AI side of things in your setup. I participated at Einstein@home for several years, not sure on their ARM application binaries in 2025.
My learning approach too…you don’t learn anything if stuff just works out of the box. Problem-solving originates from having problems in the first place.
Sounds great. Cluster is so fascinating and so much fun! I’m having a blast building my own (under construction, partly virtualized, but running)