Vertex-centric Parallel Computation of SQL Queries

TL;DR

This paper introduces TAG-join, a vertex-centric parallel algorithm for executing SQL queries that matches the efficiency of state-of-the-art join algorithms and outperforms traditional RDBMSs and Spark SQL in various settings.

Contribution

It presents a novel vertex-centric scheme for SQL query execution that is efficient, scalable, and competitive with existing database systems.

Findings

TAG-join matches the complexity of top join algorithms.

On a single server, TAG-join outperforms traditional RDBMSs on TPC benchmarks.

In distributed settings, TAG-join surpasses Spark SQL.

Abstract

We present a scheme for parallel execution of SQL queries on top of any vertex-centric BSP graph processing engine. The scheme comprises a graph encoding of relational instances and a vertex program specification of our algorithm called TAG-join, which matches the theoretical communication and computation complexity of state-of-the-art join algorithms. When run on top of the vertex-centric TigerGraph database engine on a single multi-core server, TAG-join exploits thread parallelism and is competitive with (and often outperforms) reference RDBMSs on the TPC benchmarks they are traditionally tuned for. In a distributed cluster, TAG-join outperforms the popular Spark SQL engine.