Task generation is run inline in every build, and the resulting task splits are based on recent historical task timing data. However, the resulting splits are not persisted and associated with commits or commit ranges. One consequence of this is that if you check out an older commit and submit a patch build above it, the task generators may generate different task splits than the base commit did (or would have). Another consequence is the introduction new failure modes, where task generation itself fails, which generally renders a patch build invalid. It would be better if task splits were deterministic/reproducible at each commit, and ideally precomputed: then they could not fail and no time would needed to be spent in the build to obtain them.