III: Small: Novel Paradigms for Automated Index Tuning

点击翻译按钮获取中文摘要

Indexes impact crucially the performance of a database system, and hence creating the right indexes for a workload, also known as index tuning, is an important task in database administration. Automated index tuning techniques have thus become an indispensable tool for administrators. However, existing techniques target the simplified scenario of a single database hosted on a single machine, which does not match the more complex system architectures observed in practice, e.g., multi-tenant systems with virtualized databases, or shared-nothing parallel databases. This mismatch results in suboptimal tuning that underutilizes the available system resources. Moreover, the majority of techniques require the administrator to have detailed knowledge of the anticipated database workload, which is not feasible when the workload has unpredictable characteristics (e.g., workload in ad-hoc data analysis). An alternative is techniques that analyze the workload online, but such techniques assume total control of index-maintenance decisions and essentially sidestep the administrator. The proposed project develops novel index tuning techniques that address the aforementioned shortcomings. The first contribution is a new tuning paradigm that couples online workload analysis with feedback and interaction from the administrator, thus combining the best features of previous approaches. The theoretical foundation is a novel extension of previous results from the field of online optimization. Subsequently, the paradigm is specialized for databases on compute clusters and virtualized databases, two practical architectures that are not covered by previous techniques. The project will impact the education of Computer Science students in database systems and will lead to more effective tools for database administration. For further information see the project page at http://www.cs.ucsc.edu/~alkis/tuning.