The EDBT Best paper award is given to paper #60 "Efficiently Managing Deep Learning Models in a Distributed Environment" by Nils Straßenburg (Hasso Plattner Institut)*; Ilin Tolovski (Hasso Plattner Institute); Tilmann Rabl (HPI, University of Potsdam). It will be presented in EDBT 1 on Wednesday, March 30.
As deep learning pervades into our daily lives, there is a growing interest to recover or reproduce learned models as well as to keep provenance information about those models. In their paper “Efficiently Managing Deep Learning Models in a Distributed Environment,” the authors present different mechanisms to manage learned models; they analyze the strengths and weaknesses of each of those mechanisms; and they provide the open-source tool ‘MMlibrary’ to address model management. As such, the paper contributes an important building block, not only to efficiently manage deep learning models, but also to work toward “explainable” AI systems, which is one of the key challenges in the field.
The EDBT Best paper runner-up award is given to paper #6 "Revisiting Runtime Dynamic Optimization for Join Queries in Big Data Management Systems" by Christina Pavlopoulou (University of California, Riverside)*; Vassilis J. Tsotras (UC Riverside); Michael Carey (UC Irvine). It will be presented in EDBT 4 on Wednesday, March 30.
This paper considers an old problem and an early solution within a very modern context. The topic of optimizing declarative queries has been studied since the early years of relational database systems. One approach that was proposed very early is runtime dynamic optimization, although most commercial systems display a preference for static cost-based optimization. This paper revives the idea of dynamic runtime query optimization within the context of a shared-nothing distributed big data system. The proposal eliminates the negative impact of errors in estimates in cost-based optimizers, which are exacerbated in big data systems due to the increased data size. It borrows ideas from cost-based optimizers by collecting statistics during materialization to assist in dynamic join order selection. The proposed approach is implemented within AsterixDB and demonstrates superior performance.