Reasoning on data partitioning for single-round multi-join evaluation in massively parallel systems

Abstract

Evaluating queries over massive amounts of data is a major challenge in the big data era. Modern massively parallel systems, such as, Spark, organize query answering as a sequence of rounds each consisting of a distinct communication phase followed by a computation phase. The communication phase redistributes data over the available servers, while in the subsequent computation phase each server performs the actual computation on its local data. There is a growing interest in single-round algorithms for evaluating multiway joins where data is first reshuffled over the servers and then evaluated in a parallel but communication-free way. As the amount of communication induced by a reshuffling of the data is a dominating cost in such systems, we introduce a framework for reasoning about data partitioning to detect when we can avoid the data reshuffling step. Specifically, we formalize the decision problems parallel-correctness and transfer of parallel-correctness, provide semantical characterizations, and obtain tight complexity bounds.