DKWS: A Distributed System for Keyword Search on Massive Graphs (Complete Version)

TL;DR

This paper introduces DKWS, a novel distributed system for keyword search on large graphs, featuring a monotonic search algorithm, a notify-push paradigm, and a programming model that significantly improves speed and reduces communication costs.

Contribution

The paper presents a new distributed keyword search system with a monotonic algorithm, a notify-push paradigm, and a programming model, enhancing efficiency and scalability over existing methods.

Findings

Up to 100x faster than existing techniques.

Communication costs reduced by a factor of 7.6.

Effective pruning bounds improve search efficiency.

Abstract

Due to the unstructuredness and the lack of schemas of graphs, such as knowledge graphs, social networks, and RDF graphs, keyword search for querying such graphs has been proposed. As graphs have become voluminous, large-scale distributed processing has attracted much interest from the database research community. While there have been several distributed systems, distributed querying techniques for keyword search are still limited. This paper proposes a novel distributed keyword search system called $\DKWS$. First, we \revise{present} a {\em monotonic} property with keyword search algorithms that guarantees correct parallelization. Second, we present a keyword search algorithm as monotonic backward and forward search phases. Moreover, we propose new tight bounds for pruning nodes being searched. Third, we propose a {\em notify-push} paradigm and $\PINE$ {\em programming model} of $\DKWS$. The notify-push paradigm allows {\em asynchronously} exchanging the upper bounds of matches across the workers and the coordinator in $\DKWS$. The $\PINE$ programming model naturally fits keyword search algorithms, as they have distinguished phases, to allow {\em preemptive} searches to mitigate staleness in a distributed system. Finally, we investigate the performance and effectiveness of $\DKWS$ through experiments using real-world datasets. We find that $\DKWS$ is up to two orders of magnitude faster than related techniques, and its communication costs are $7.6$ times smaller than those of other techniques.