Optimal parameters for bloom-filtered joins in Spark

Abstract: In this paper, we present an algorithm that joins relational database tables efficiently in a distributed environment using Bloom filters of an optimal size. We propose not to use fixed-size bloom filters as in previous research, but to find an optimal size for the bloom filters, by creating a mathematical model of the join algorithm, and then finding the optimal parameters using traditional mathematical optimization. This algorithm with optimal parameters beats both previous approaches using bloom filters and the default SparkSQL engine not only on star-joins, but also on traditional database schema. The experiments were conducted on a standard TPC-H database stored as parquet files on a distributed file system.