Reading a crash cluster without drowning in volume

Raw count is a poor ranking. A cluster that fires on every cold start can look huge and still be an old, accepted native library. A cluster that fires only after resume on one Android version can look small and still be this week’s store mail.

We start with a paper timeline of the last two version codes, then filter the console to those codes only. Anything that existed at the same rate two versions ago is labelled aged unless the team has a reason to reopen it. The remaining clusters are grouped by entry path: launch, resume, checkout, map, playback.

Three is an arbitrary ceiling we still use. If four clusters are genuinely new, the brief says four. If product wants a single root cause and the traces disagree, the brief says they disagree. Crash and stability analytics that pretends otherwise wastes the next stand-up.