View article

While estimating the result size of a group-by operation on a base table is hard on its own, the presence of selections makes this problem increasingly difficult to solve. We show that skewed data distributions and correlations found in real-world data heavily affect the results of traditional cardinality estimators. On the other hand, deep learning has recently been shown to be a more robust approach to cardinality estimation. Our evaluation shows that our (setbased) deep learning model significantly enhances the quality of filtered group-by cardinality estimates.