The little ARM board that drives my garage sensors had been rebooting at random for weeks. Always under load, never idle, which is the sort of symptom that sends you chasing software ghosts for far too long. I rewrote the watchdog, swapped the SD card, even reflashed the bootloader. None of it mattered, because the fault was not in any of that. It was a hairline crack in a solder joint on the power input, and the only fix was an iron and a steady-ish hand.
The giveaway, eventually, was that I could make it reboot by pressing gently near the barrel jack. Tap the corner of the PCB, watch the kernel panic. That is not software. That is mechanics. A cold or cracked joint behaves exactly like an intermittent connection because that is precisely what it is, and the more current you pull through it the more the tiny resistance heats and shifts and breaks contact.
Under a loupe the joint looked dull and grey rather than shiny, with a faint ring around the pad where it had been flexing. Classic. The board had clearly shipped with a slightly dry joint and a couple of years of thermal cycling had finished the job.
The repair itself was almost anticlimactic once I knew what I was doing:
- Clean the old joint with a little fresh flux. Old, oxidised solder does not want to flow, and dragging a dry iron across it just makes a mess.
- Add a small amount of fresh leaded solder. I keep a reel of 60/40 around precisely because it flows at a sane temperature and behaves predictably. Lead-free is the right thing environmentally but it is unkind to amateurs.
- Heat the pad and the pin together, not just the blob, until everything wets out and goes shiny.
- Let it cool without breathing on it or nudging the board. A disturbed joint while it solidifies is how you create the next dry joint.
Iron set to about 350°C, a chisel tip rather than the needle-fine one because you want thermal mass to carry heat into the joint quickly and get out again. The longer you sit there cooking the pad, the more you risk lifting it off the board entirely, and a lifted pad is a genuinely annoying repair.
Total time at the bench, maybe ten minutes including the cup of tea. The board has now run for four days under the same load that used to topple it inside an hour, and I have stopped pressing on the corner hopefully every time I walk past.
There is a small lesson buried in here that I keep relearning. When a system fails under load but not at idle, and no amount of software poking changes the shape of the failure, start suspecting the physical layer. Power, connectors, joints, the bits we stop thinking about the moment they work once. An iron, some flux and a magnifier saved me a tenner on a replacement and, more to the point, an evening of further software archaeology that would have found nothing.