Let me save you the suspense. A four-node Raspberry Pi cluster will not teach you anything you couldn't learn faster on a single machine with some VMs and a bit of imagination. I knew that going in. I built it anyway, and I had a thoroughly good time, and I refuse to feel bad about it.
The premise was simple enough. Four Pi 3 Model B boards, a small unmanaged switch, a stack of acrylic plates with brass standoffs, and one of those five-port USB chargers that promises 2.4A a port and delivers something closer to "enough, probably". The plan was to run a little Kubernetes on it, or k3s once that lands, or just Docker Swarm, and feel like I had a data centre on my desk. What I actually got was a small heater that occasionally served a web page.
The first lesson, if you can call it that, was about power. The Pi is fussy. Undervolt it even slightly and you get the little lightning bolt in the corner and a cascade of weird, unreproducible faults: SD cards that corrupt for no reason, USB devices that drop, nodes that reboot under load. I spent a genuinely embarrassing evening debugging a "network problem" that was just node three browning out whenever I pushed an image to it. The fix was better cables and a proper supply, not anything clever. That is the recurring theme of small ARM clusters: the interesting failures are almost always boring power and SD card problems wearing a costume.
The second lesson was that the cluster was slower than my laptop at everything that mattered, and that this was fine, because speed was never the point. Four Cortex-A53 cores per node sounds like sixteen cores until you remember they are 1.2GHz and sipping power and that the network between them is 100Mbit because the Pi 3's ethernet hangs off the USB bus. You do not build this to get work done. You build it because there is something deeply satisfying about four little green lights blinking in unison.
What it was genuinely good for was making distributed-systems failures physical. When you kill a node by yanking its power, and you can watch the others notice, that lands differently than draining a VM. I learned more about how a scheduler reschedules pods by unplugging things with my actual hand than I ever did reading the docs. There is a tactile honesty to it. The cluster does not pretend a node is healthy when you are holding its power lead in the other hand.
So, was it useful? No. Would I recommend it? Absolutely, with the caveat that you should know what you are buying. You are not buying a cheap server. You are buying a toy that teaches patience, a renewed respect for power budgets, and the particular joy of a blinking LED stack on your desk. I keep mine running. It does almost nothing. I would not get rid of it for the world.