Some areas for improvement are:

• minimizing code duplication, e.g. creation of the SamplingEstimatorImpl instances
• make the document generation process more robust. One way we could achieve this would be to allow the caller to specify document schemas of the collections we are testing
• fix typos, e.g. iniitalizeSamplingAlgoBasedOnChunk()
• use parse and fromjson to construct the MatchExpressions