I offer you a thought experiment. For your current project you've set up some inter-process communication. Nothing tricky involved, just your standard client–server deal. You've even outsourced the protocol design and are using Google's code to handle all the grody details of serializing and deserializing. Well, okay, on the server side you're using someone else's code but it implements the same protocol, right? Now you run into a bug.
The vast majority of the time everything works smoothly— even verified by taking SHA1 hashes of the messages on both sides and comparing them. But every so often the Java client crashes. In particular, it crashes whenever reading a result message (from the server) of length 255 or 383 and maybe some larger sizes. It does, however, work perfectly fine for intervening message lengths (including 254 and 256). So what's wrong?
Knowing the answer, you'd predict difficulties with length 511 as well, though you haven't observed it to fail (or succeed) in practice.
We're using the "delimited" version of protocol buffers, the one that writes the message/payload length just before the payload.
It's a one-line bugfix.