CRII: III: RUI: Adaptive Query Processing for Crowd-Powered Database Systems

点击翻译按钮获取中文摘要

Database systems provide users with the ability to ask questions, or queries, about collections of data that the system stores (e.g., find employees who had worked in the company for at least 2 years) and provide the answers very fast. People are better equipped than computers to tackle problems that require judgement or data interpretation due to their real-world experience and perception. A crowd-powered database system uses groups of people called "the crowd" to help with answering users' queries by recruiting them to process data using criteria that are subjective and/or require visual or semantic interpretation. For example, a user may want to find a set of faculty job postings in which the job description discusses a commitment to diversity and for which the school is in a safe location; interpreting each job description and researching crime statistics are tasks well-suited for people to perform. The system can coordinate crowd workers to process data more efficiently than the user alone could, which is advantageous when there are more than a handful of data items to process. While queries processed by the crowd may take hours or days to complete, crowd-powered database systems enable the processing of complex queries. For example, queries such as determining which research articles about a certain medical device contain experimental results comparing this and other devices, or finding out which of a set of jewelers only use ethically sourced metals and stones and also ship to Alaska. Database systems are designed to optimize the efficiency of query processing of individual users. A query often involves multiple parts, e.g., for the job postings query these are (1) filter out jobs that do not describe a commitment to diversity and (2) filter out jobs for schools in an unsafe location. A job that does not meet the first criterion does not need to be processed for the second one, and vice versa. The processing order for the parts of the query influences how much computation is needed and how long the query will take to process. Traditional database systems have information about how long parts of a query will take and the likelihood of items satisfying filters; they use this information to choose an efficient processing ordering for a query. However, this information is not known for crowd-powered database systems. The usefulness of optimizers for crowd-powered database systems hinges on their ability to find an efficient way to process a user's query when this information is unknown before processing the query. The aim of this research project is to tackle this challenge by developing a system to process queries involving multiple filtering criteria that observes the execution environment and adjusts its processing strategy as the query executes. This project will have broad impact by yielding a query processing system that will empower users to ask more interesting questions about data, advancing research in allocating human computation resources in dynamic environments, as well as training a group of undergraduate students both in research and in the principles of systems design.The goal of this research is to build a cost-based query optimizer for crowd-powered filter queries for which important statistics used in optimization are unknown at query time. These statistics include traditional metrics such as filter selectivity as well as new contributors to query cost such as the time it takes crowd workers to complete a unit of work and the number of workers needed to reach consensus for a subjective evaluation. The project takes an adaptive approach to query processing: while the query is running, the system observes cost and selectivity information and periodically reorders the query plan operators to reduce overall query cost. The researchers will demonstrate that their query optimizer yields query costs that are comparable to costs from the optimal crowd-based query plan for which selectivity and subjectivity information is known a priori. Source code, papers, and presentations are available on the project web site (https://www.cs.hmc.edu/~beth/adaptivecrowd.shtml).