I have a confession: the four-node Pi cluster on my shelf does nothing useful. It runs a k3s control plane and three workers, it serves exactly one toy web app, and the whole arrangement could be replaced by a single process on the NAS I already own. I built it anyway, and I'd do it again.
The honest appeal was the wiring. A 3D-printed stack, four Pi 4s, a flat run of short Ethernet patch leads into a cheap unmanaged switch, and a single USB-C power brick that I spent far too long convincing myself wouldn't sag under load. The soldering iron only came out for a janky little fan header, which promptly buzzed at a frequency precisely tuned to annoy me from across the room.
What I actually learned was nothing I can put on a CV. I already knew how Kubernetes schedules pods. I already knew etcd hates flaky storage, and that SD cards are flaky storage, so I booted from USB SSDs and felt very clever for an afternoon. None of it was a revelation. All of it was satisfying in the way that tidying a cupboard is satisfying: low stakes, visible progress, a thing that is better arranged at the end than it was at the start.
There's a quiet argument for building things that don't matter. When the cluster falls over, nobody is paged, nobody loses money, and I get to poke at it with genuine curiosity instead of dread. That's a rare mood in this job, and I'll take it where I can find it. The Pis blink away on the shelf, useless and content, and so am I.