Continuously Providing Approximate Results under Limited Resources: Load Shedding and Spilling in XML Streams

Because of the high volume and unpredictable arrival rates, stream processing systems may not always be able to keep up with the input data streams, resulting in buffer overflow and uncontrolled loss of data. To continuously supply online results, two alternate solutions to tackle this problem of unpredictable failures of such overloaded systems can be identified. One technique, called load shedding, drops some fractions of data from the input stream to reduce the memory and CPU requirements of the workload. However, dropping some portions of the input data means that the accuracy of the output is reduced since some data is lost. To produce eventually complete results, the second technique, called data spilling, pushes some fractions of data to persistent storage temporarily when the processing speed cannot keep up with the arrival rate. The processing of the disk resident data is then postponed until a later time when system resources become available. This dissertation explores these load reduction technologies in the context of XML stream systems. Load shedding in the specific context of XML streams poses several unique opportunities and challenges. Since XML data is hierarchical, subelements, extracted from different positions of the XML tree structure, may vary in their importance. Further, dropping different subelements may vary in their savings of storage and computation. Hence, unlike prior work in the literature that drops data completely or not at all, in this dissertation we introduce the notion of structure-oriented load shedding, meaning selectively some XML subelements are shed from the possibly complex XML objects in the XML stream. First we develop a preference model that enables users to specify the relative importance of preserving different subelements within the XML result structure. This transforms shedding into the problem of rewriting the user query into shed queries that return approximate answers with their utility as measured by the user preference model. Our optimizer finds the appropriate shed queries to maximize the output utility driven by our structure-based preference model under the limitation of available computation resources. The experimental results demonstrate that our proposed XML-specific shedding solution consistently achieves higher utility results compared to the existing relational shedding techniques. Second, we introduces structure-based spilling, a spilling technique customized for XML streams by considering the spilling of partial substructures of possibly complex XML elements. Several new challenges caused by structure-based spilling are addressed. When a path is spilled, multiple other paths may be affected. We categorize varying types of spilling side effects on the query caused by spilling. How to execute the reduced query to produce the correct runtime output is also studied. Three optimization strategies are developed to select the reduced query that maximizes the output quality. We also examine the clean-up stage to guarantee that an entire result set is eventually generated by producing supplementary results to complement the partial results output earlier. The experimental study demonstrates that our proposed solutions consistently achieve higher quality results compared to the state-of-the-art techniques. Third, we design an integrated framework that combines both shedding and spilling policies into one comprehensive methodology. Decisions on the choice of whether to shed or spill data may be affected by the application needs and data arrival patterns. For some input data, it may be worth to flush it to disk if a delayed output of its result will be important, while other data would best directly dropped from the system given that a delayed delivery of these results would no longer be meaningful to the application. Therefore we need sophisticated technologies capable of deploying both shedding and spilling techniques within one integrated strategy with the ability to deliver the most appropriate decision customers need for each specific circumstance. We propose a novel