Detecting coherent explorations in SQL workloads. (September 2020)

• Record Type: Journal Article
• Authors: Peralta, Verónika; Marcel, Patrick; Verdeaux, Willeme; Diakhaby, Aboubakar Sidikhy
• Is Part Of: Information systems. Volume 92(2020)
• Publication Date: 2020-09
• Subjects: Database management -- Periodicals; Electronic data processing -- Periodicals; Bases de données -- Gestion -- Périodiques; Informatique -- Périodiques; Database management; Electronic data processing; Periodicals; 005.7
• DOI: 10.1016/j.is.2019.101479
• Languages: English
• ISSNs: 0306-4379

Abstract

Abstract: This paper presents a proposal aiming at better understanding a workload of SQL queries and detecting coherent explorations hidden within the workload. In particular, our work investigates SQLShare (Jain et al., 2016), a database-as-a-service platform targeting scientists and data scientists with minimal database experience, whose workload was made available to the research community. According to the authors of (Jain et al., 2016), this workload is the only one containing primarily ad-hoc hand-written queries over user-uploaded datasets. We analyzed this workload by extracting features that characterize SQL queries and we investigate three different machine learning approaches to use these features to separate sequences of SQL queries into meaningful explorations. The first approach is unsupervised and based only on similarity between contiguous queries. The second approach uses transfer learning to apply a model trained over a dataset where ground truth is available. The last approach uses weak labeling to predict the most probable segmentation from heuristics meant to label a training set. We ran several tests over various query workloads to evaluate and compare the proposed methods. Highlights: We propose an approach for segmenting a SQL workload into meaningful, coherent database explorations. We describe each query of a SQL workload with a set of features that are intrinsic to a query or that relate the query to its neighbor in the workload. We define a set of similarity indexes, exploiting query features, that quantify the similarity between contiguous queries from several points of view. We proposed three methods for workload segmentation: unsupervised learning, supervised learning and weak supervision.