Currently, the --fill-missing heuristic can be described as below:

If results are missing on a bot,
1. Find another bot whose OS is the same as the missing bot (the version is likely different).
2. If none is found, choose a random bot.
Then the chosen bot is used to fill in the missing results.

We could be a bit smarter, taking into account the baseline fallback path. For example, prefer to fill in missing Linux results with Windows results, Android results with Linux results, etc.


Besides, the current fill-missing logic works on the level of bots (i.e. only works if the whole result is missing on a bot). We might consider changing it to work on a per-test level. This will cover more use cases, but I haven't thought too much about the idea.