POLYTOPE: a flexible sampling system for answering exploratory queries

摘要

Data exploration task is usually quite time-consuming. Analysts who want to find interests or verify their hypothesis may prefer a lower response time while tolerating a bounded error. Approximate query processing (AQP) is a convincing way to achieve this goal by leveraging some pre-computed samples to speed up this process. Existing sampling based AQP systems usually take a single sampling strategy on the whole dataset. However, during the data exploration tasks, various potential interests may distribute in different parts of dataset. To explore these interests, queries submitted by users thus show a rich diversity for separate sub-datasets. Therefore, only one single sampling strategy is obviously not competent for all queries accessing various sub-datasets. In this paper, we proposed a flexible and effective sampling system POLYTOPE especially designed for the data exploration tasks. To achieve this, we take the following three key ideas: (1) split the dataset into sampling blocks according to the user query patterns, (2) individually generate a set of optimized samples for each sampling block, and (3) automatically select an optimal sample at run time. We utilize both user query patterns and underlying data distribution to fulfill these ideas. We have implemented our system on the Spark platform and our comprehensive experimental results show that our system improved the accuracy performance up to 46% under the same time constraint for the data exploration tasks.