ABSTRACT:

Despite advances in computing power, the cost of large-scale analytics and machine learning remains daunting to small and large enterprises alike. This has created a pressing demand for reducing infrastructure costs and query latencies. To meet these goals, data analysts and applications are in many cases willing to tolerate a slight — but controlled — degradation of accuracy in exchange for substantial gains in cost and performance, which we refer to as statistical trade-offs. This is particularly true in the early stages of data exploration and is in stark contrast to traditional trade-offs where the infrastructure costs must increase for higher performance.

My research builds large-scale data systems that can make these statistical trade-offs in a principled manner. In this talk, I will focus on two specific directions. First, I will present VerdictDB, a system that enables quality-guaranteed, statistical trade-offs without any changes to backend infrastructure; thus, it offers a universal solution for off-the-shelf query engines. Second, I will introduce Database Learning, a new query execution paradigm that allows existing query engines to constantly learn from their past executions and become “smarter” over time without any user intervention. I will conclude by briefly discussing other promising directions with emerging workloads beyond SQL, including visualization and machine learning.

BIO:

 Yongjoo Park is a research fellow in computer science and engineering at the University of Michigan, Ann Arbor. His research interest is software systems for fast data analytics and machine learning. He received a Ph.D. from the University of Michigan, advised by Michael Cafarella and Barzan Mozafari. He is a recipient of 2018 ACM SIGMOD Jim Gray Dissertation Award Runner-up, Kwanjeong Ph.D. Fellowship, and Jeongsong Graduate Study Fellowship.