Abstract
A database session is a sequence of requests presented to the database system by a user or an application to achieve a certain task. Session identification is an important step in discovering useful patterns from database trace logs. The discovered patterns can be used to improve the performance of database systems by prefetching predicted queries, rewriting the current query or conducting effective cache replacement.
In this paper, we present an application of a new session identification method based on statistical language modeling to database trace logs. Several problems of the language modeling based method are revealed in the application, which include how to select values for the parameters of the language model, how to evaluate the accuracy of the session identification result and how to learn a language model without well-labeled training data. All of these issues are important in the successful application of the language modeling based method for session identification. We propose solutions to these open issues. In particular, new methods for determining an entropy threshold and the order of the language model are proposed. New performance measures are presented to better evaluate the accuracy of the identified sessions. Furthermore, three types of learning methods, namely, learning from labeled data, learning from semi-labeled data and learning from unlabeled data, are introduced to learn language models from different types of training data. Finally, we report experimental results that show the effectiveness of the language model based method for identifying sessions from the trace logs of an OLTP database application and the TPC-C Benchmark.
