Quality-aware adaptation in database systems

This dissertation addresses the problem of quality adaptation in databases. We focus on system support and various related techniques in two types of database systems: multimedia databases and stream databases. We are specifically interested in how to fully utilize system resources such that system-level performance targets (e.g., throughput, user satisfaction) can be maximized. We identify specific challenges and propose different solutions in these two types of databases. In multimedia databases, users can request data with a specific quality requirement due to the needs of applications or the limitations of client-side resources. A widely-used method to support heterogeneous user quality needs is to pre-compute and store multiple quality replicas of data items. However, it causes unacceptable increases in storage costs to store all possible quality-specific replicas. Therefore, how to select quality-specific replicas for media objects under an overall storage constraint becomes a problem. The fact that the popularity of each possible replica changes over time makes the problem more challenging. We develop fast approximate algorithms to solve this problem under two quality service models. The continuous feature of both data and queries in data stream management systems (DSMSs) places great demands on system resources. However, queries can be processed with different levels of quality such as timeliness, reliability, and precision. In such systems, we focus on how to control tuple processing delays, which is an important quality metric in many data stream applications. When the system is overloaded, a widely-used approach to maintain tuple delays in DSMS query processing is load shedding, i.e., dropping data. The problem of interest is to determine the appropriate time and amount of load shedding such that tuple processing delays are maintained under a desirable level with minimal data loss. We propose a solution based on feedback control theory with significantly improved processing delays and the same amount of data loss, as compared to existing solutions in popular data stream systems