The RepositionLastSplitPointsTest test is very "sensible" as it expect the auto splitter behaviour to be very precise.

It helped spotting the following bug that needs to be solved in order to avoid future failures: this strict gt comparison needs to be replaced by a >=, otherwise the generated chunks will always contain one key more than the expected size.

Moreover, the size of the string inserted in the generated documents should be changed to 1024^2 - 24 in order to obtain documents of exactly 1MB (contrary to the calculation done under SERVER-60463, that was done by relying on a wrong estimated data size).