Database Ergalics: Metrology, Suboptimality, and Thrashing
Assistant Professor, Kyungpook National University (School of Computer Science and Engineering), 2017 – current
Senior Researcher, KISTI (National Institute of Supercomputing & Networking), 2005 – 2017
Ph.D., University of Arizona (Computer Science), 2015
M.S., KAIST (Computer Science), 2005
B.S., Kyungpook National University (Computer Science), 2003
In the database field, while some very strong mathematical and engineering work has been done, a scientific perspective has been much less prominent. Taking a scientific approach can lead to a deep understanding of DBMSes’ behaviors for better engineered designs. Database ergalics, born to realize this scientific methodology, aims at better understanding DBMSes as a “general” class of computational artifacts via a scientifically rigorous approach, to come up with insights and ultimately with predictive theories about how they behave, again, as a general class.
In this talk, I will introduce our troika of research topics explored in the database ergalics area. More specifically, I will show concrete examples through empirical evidences on several phenomena observed “across” modern relational DBMSes: varying query time (i.e., metrology), suboptimal query plans (i.e., suboptimality), and falling transaction throughput (i.e., thrashing). I will then explain the causality of each of the phenomena via a structural causal model that we proposed and statistically validated on the empirical data obtained from the DBMSes. During my talk, I will present a novel DBMS-oriented research infrastructure, called AZDBLab, that contributed to the data collection, as well as (i) execution time protocol, (ii) query timing protocol, and (iii) thrashing analysis protocol that we invented for the studies. Finally, I will suggest important engineering implications based on our findings.