Initially, we added just a doc-count and date-based check to the rollover API. We avoided a size-based check because it could be wildly inaccurate if multiple merges are in progress on different shards of the index.

We could make a more reasonable estimation by converting from segment sizes to # of docs which makes us less prone to merges etc. Could we sample the avg doc size on disk from existing docs and make the decision ultimately based on docs?