Optimization of Progressive Queries via Materialized Views for Large Databases

Abstract

There is an increasing demand to efficiently process emerging types of queries, such as progressive queries (PQ), on large scale databases from numerous contemporary applications including telematics, e-commerce, and social media. Unlike a conventional query, a PQ consists of a set of interrelated step-queries (SQ). A user formulates a new SQ on the fly based on the result(s) from the previously executed SQ(s). Processing PQs raises a number of new challenges. Existing database management systems were not designed to efficiently process such queries. In this dissertation, we propose a suite of novel materialized-view based techniques to efficiently process PQs. First, we propose a dynamic materialized-view based approach to efficiently processing a special type of PQs, called monotonic linear PQs. We introduce a so-called superior relationship graph to capture superior relationships among SQs of such a PQ and suggest a method to estimate the benefit of keeping the result of an SQ as a materialized view using the graph. To efficiently construct the superior relationship graph, we propose two algorithms: generating-based and pruning-based. To improve the view searching efficiency and quality, we design an algorithm with a special storage structure to store and manage the materialized views. Second, to handle generic PQs, we define a so-called multiple query dependency graph to capture the data source dependency relationships that exist among SQs and external tables of a generic PQ. Using the graph, a mathematical benefit estimation model, which takes both the impact and the effectiveness of materialization into consideration, is derived. A greedy method and a dynamic programming method to solve the view maintenance problem are proposed. Third, to efficiently find usable materialized views from the view space/set for answering a given SQ, we suggest a dynamic materialized view index method. A special index tree structure with nodes ordered by a two-level priority rule that facilitates efficient locating of different types of nodes is designed. Bitmaps encoded with special methods are also used to refine the pruning of unusable views during a search. Fourth, to support PQs in a big data environment like Hadoop, we propose an index based technique for performing a new column family join operation on Hbase tables. To efficiently process such a join operation, we suggest a multiple freedom family index. A parallel MapReduce algorithm to construct the index is developed. To perform a column family join on two Hbase tables using the indexes, we present two partitioning methods to balance the workload among map nodes in a MapReduce algorithm. The introduced column family join operation and its relevant processing technique can ensure the closure property that is essential to the processing of PQs. To examine the performance of the proposed techniques, we performed extensive empirical and theoretical analyses. Our studies show that the proposed techniques are quite promising in efficiently processing PQs. To our knowledge, our work is the first to apply the materialized-view based approach to efficiently processing progressive queries on large databases.