Finding a Second Wind: Speeding Up Graph Traversal Queries in RDBMSs Using Column-Oriented Processing

TL;DR

This paper introduces a novel column-store based approach using late materialization to significantly accelerate recursive graph traversal queries in RDBMSs, demonstrating up to 6x speedups over PostgreSQL.

Contribution

It presents a new method leveraging position-aware column-store technology and Volcano-style operators to improve recursive query performance in RDBMSs.

Findings

Up to 6x faster breadth-first search in PosDB compared to PostgreSQL.

Position-based approach outperforms tuple-based approach in recursive queries.

Column-stores with position-awareness cannot be efficiently emulated by row-stores.

Abstract

Recursive queries and recursive derived tables constitute an important part of the SQL standard. Their efficient processing is important for many real-life applications that rely on graph or hierarchy traversal. Position-enabled column-stores offer a novel opportunity to improve run times for this type of queries. Such systems allow the engine to explicitly use data positions (row ids) inside its core and thus, enable novel efficient implementations of query plan operators. In this paper, we present an approach that significantly speeds up recursive query processing inside RDBMSes. Its core idea is to employ a particular aspect of column-store technology (late materialization) which enables the query engine to manipulate data positions during query execution. Based on it, we propose two sets of Volcano-style operators intended to process different query cases. In order validate our ideas, we have implemented the proposed approach in PosDB, an RDBMS column-store with SQL support. We experimentally demonstrate the viability of our approach by providing a comparison with PostgreSQL. Experiments show that for breadth-first search: 1) our position-based approach yields up to 6x better results than PostgreSQL, 2) our tuple-based one results in only 3x improvement when using a special rewriting technique, but it can work in a larger number of cases, and 3) both approaches can't be emulated in row-stores efficiently.