Modeling Large Time Series for Efficient Approximate Query Processing

Evolving customer requirements and increasing competition force business organizations to store increasing amounts of data and query them for information at any given time. Due to the current growth of data volumes, timely extraction of relevant information becomes more and more difficult with traditional methods. In addition, contemporary Decision Support Systems (DSS) favor faster approximations over slower exact results. Generally speaking, processes that require exchange of data become inefficient when connection bandwidth does not increase as fast as the volume of data. In order to tackle these issues, compression techniques have been introduced in many areas of data processing. In this paper, we outline a new system that does not query complete datasets but instead utilizes models to extract the requested information. For time series data we use Fourier and Cosine transformations and piece-wise aggregation to derive the models. These models are initially created from the original data and are kept in the database along with it. Subsequent queries are answered using the stored models rather than scanning and processing the original datasets. In order to support model query processing, we maintain query statistics derived from experiments and when running the system. Our approach can also reduce communication load by exchanging models instead of data. To allow seamless integration of model-based querying into traditional data warehouses, we introduce a SQL compatible query terminology. Our experiments show that querying models is up to 80% faster than querying over the raw data while retaining a high accuracy.