YSU_Yu

Approximate query processing (or AQP) is an emerging research topic in big data analytics. AQP focuses on deriving fast and accurate estimations for complex queries that are usually time-consuming and expensive to run on large datasets. Traditional methods, such as histogram and sketch, are insufficient when applied to big data because of the processing limits. An essential question lacking research is how to assess the errors of AQP estimations. This project focuses on assessing the errors of AQP query estimations, especially for common selection queries. Traditional methods can generate confidence intervals for query estimations based on strict assumptions such as the normal distribution assumption. Therefore, they are not applicable to massive datasets. In this project, the PI will employ a novel non-parametric statistical method called bootstrap sampling which requires less strict assumptions and brings many statistical advantages. A prototype system will be developed employing bootstrap sampling to efficiently compute standard errors and confidence intervals for AQP systems, especially those answering selection queries, namely σ-AQP. Selection queries comprise a large portion of daily data queries. For broader applications, this framework will allow selection queries to include common aggregation operators such as average, sum, and count. The PI will investigate the computing and storage costs when bootstrap replicas are computed. A framework will be developed to automate both the AQP estimation and error estimation operations. Extensive benchmarks will be performed on large datasets such as the TPC-H benchmark.