The datasets are pretty small (<=5k docs), which means the indexes are pretty small. With SERVER-103623, we should be able to get pretty good accuracy for index-based CE. From my experimenting, we will also either need to extend index-based CE to handle more kinds of index-bounds (SERVER-103624 e.g. bounds that are not a single, simple interval) or we will need to disable some of the test cases.

For sampling CE, we may run into the issue where very selective predicates produce inaccurate results. It's unclear if we'll need to adjust sampling settings to get better accuracy