2019 Theses Doctoral
Physical Plan Instrumentation in Databases: Mechanisms and Applications
Database management systems (DBMSs) are designed with the goal set to compile SQL queries to physical plans that, when executed, provide results to the SQL queries. Building on this functionality, an ever-increasing number of application domains (e.g., provenance management, online query optimization, physical database design, interactive data profiling, monitoring, and interactive data visualization) seek to operate on how queries are executed by the DBMS for a wide variety of purposes ranging from debugging and data explanation to optimization and monitoring. Unfortunately, DBMSs provide little, if any, support to facilitate the development of this class of important application domains. The effect is such that database application developers and database system architects either rewrite the database internals in ad-hoc ways; work around the SQL interface, if possible, with inevitable performance penalties; or even build new databases from scratch only to express and optimize their domain-specific application logic over how queries are executed.
To address this problem in a principled manner in this dissertation, we introduce a prototype DBMS, namely, Smoke, that exposes instrumentation mechanisms in the form of a framework to allow external applications to manipulate physical plans. Intuitively, a physical plan is the underlying representation that DBMSs use to encode how a SQL query will be executed, and providing instrumentation mechanisms at this representation level allows applications to express and optimize their logic on how queries are executed.
Having such an instrumentation-enabled DBMS in-place, we then consider how to express and optimize applications that rely their logic on how queries are executed. To best demonstrate the expressive and optimization power of instrumentation-enabled DBMSs, we express and optimize applications across several important domains including provenance management, interactive data visualization, interactive data profiling, physical database design, online query optimization, and query discovery. Expressivity-wise, we show that Smoke can express known techniques, introduce novel semantics on known techniques, and introduce new techniques across domains. Performance-wise, we show case-by-case that Smoke is on par with or up-to several orders of magnitudes faster than state-of-the-art imperative and declarative implementations of important applications across domains.
As such, we believe our contributions provide evidence and form the basis towards a class of instrumentation-enabled DBMSs with the goal set to express and optimize applications across important domains with core logic over how queries are executed by DBMSs.
Files
- PSALLIDAS_columbia_0054D_15243.pdf application/pdf 8.27 MB Download File
More About This Work
- Academic Units
- Computer Science
- Thesis Advisors
- Wu, Eugene
- Degree
- Ph.D., Columbia University
- Published Here
- June 3, 2019