We should have tests for histogram creation that test:

• mixed-type histograms (all types/ different type combinations) with different numbers of buckets
• histograms that generate buckets with range values
• histograms generated on a very large number of values
• array and scalar variants of the above testing scenarios

As a note, work done in SERVER-71310 may prove useful in ensuring that valid histograms are generated in all cases by maxdiff, but we may also want to hand-craft some small datasets that should generate a specific shape of maxdiff histogram and unit-test that this is in fact what we generate.