Most kernel panics I've chased have the decency to be intermittent, which is to say impossible. This one was a gift: it fell over every single time, given enough traffic. One box out of a dozen identical ones, and it would last about forty minutes under real load before the console filled with a stack trace and the whole thing wedged.
The trace pointed into the NIC driver, somewhere in the receive path. Same kernel, same firmware, same everything as its siblings, so my first instinct was hardware. It usually is, and it's usually wrong.
What made it reproducible was load. Below a certain packets-per-second it was fine for days. Above it, the panic arrived like clockwork. I could trigger it on demand with iperf3 and a couple of parallel streams, which after weeks of "can't reproduce" tickets felt almost luxurious. A bug you can summon is a bug you can kill.
The cause turned out to be a single duff card with a slightly different revision than the rest of the batch, paired with a driver that only exercised the broken path once the ring buffer filled. Swapped the card, ran the same iperf3 for two hours, no panic. I kept the failing one on my desk for a while as a reminder that "identical hardware" is a thing people say, not a thing that exists.