Optimal Query Plans for Geo-distributed Data Analytics at Scale

Geo-distributed Data Analytics (GDA) is crucial for organizations handling global data, utilizing numerous data centers (DCs) worldwide. Prior works look to optimize geo-distributed queries by moving the computation closer to the data, and/or to decide join order and job location heuristically. This led to generate low-quality GDA plans. In this work, we propose a novel approach to holistically optimize join order and location, based on dynamic programming, to produce optimal GDA plans. This is built on a new cost model incorporating additional GDA-parameters such as WAN cost, DC locations, heterogeneous DC capabilities. Our strong search space pruning technique helps us scale to hundreds of DCs with small overheads while retaining plan optimality. We implement our solution, GDA-OPT, on an open-source big data system that supports cross-DC analytics. The results show plan improvement (execution performance) over the state-of-the-art solution by 4X on average.