Please note: This PhD defence will take place online.
Brad
Glasbergen,
PhD
candidate
David
R.
Cheriton
School
of
Computer
Science
Supervisor: Professor Khuzaima Daudjee
Database systems are ubiquitous; they serve as the cornerstone of modern application infrastructure due to their efficient data access and storage. Database systems are commonly deployed in a wide range of environments, from transaction processing to analytics.
Unfortunately, this broad support comes with a trade-off in system complexity. Database systems contain many components and features that must work together to meet client demand. Administrators responsible for maintaining database systems face a daunting task: they must determine the access characteristics of the client workload they are serving and tailor the system to optimize for it. Complicating matters, client workloads are known to shift in access patterns and load. Thus, administrators continuously perform this optimization task, refining system design and configuration to meet ever-changing client request patterns.
Researchers have focused on creating next-generation, natively adaptive database systems to address this administrator burden. Natively adaptive database systems construct client-request models, determine workload characteristics, and tailor processing strategies to optimize accordingly. These systems continuously refine their models, ensuring they are responsive to workload shifts. While these new systems show promise in adapting system behaviour to their environment, existing, popularly-used database systems lack these adaptive capabilities. Porting the ideas in these new adaptive systems to existing infrastructure requires monumental engineering effort, slowing their adoption and leaving users stranded with their existing, non-adaptive database systems.
In this thesis, I present Dendrite, a framework that easily “bolts on” to existing database systems to endow them with adaptive capabilities. Dendrite captures database system behaviour in a system-agnostic fashion, ensuring that its techniques are generalizable. It compares captured behaviour to determine how system behaviour changes over time and with respect to idealized system performance. These differences are matched against configurable adaption rules, which deploy user-defined functions to remedy performance problems. As such, Dendrite can deploy whatever adaptions are necessary to address a behaviour shift and tailor the system to the workload at hand. Dendrite has low tracking overhead, making it practical for intensive database system deployments.
To receive the Zoom link for this PhD defence, please e-mail Professor Khuzaima Daudjee.