Labeled Property Graphs: SQL or NoSQL?

Abstract

There are two main approaches to graph databases: based on RDF model and based on labeled property graph model. RDF is well known and studied, but modern graph databases with labeled property graph model are studied much lesser. In this paper we evaluated several possible solutions for storing and querying graph data using Gremlin - general purpose graph query language from Apache TinkerPop. We used LDBC Graphalytics framework and compared NoSQL-based setups with SQL-based setups. We evaluated JanusGraph on HBase both on single machine and cluster and SQLG on top of PostgreSQL and H2. We used datasets from the different domains and of different sizes up to tens of millions vertices and edges. Evaluation results show that for the used workload SQLG with PostgreSQL is about ten times faster than JanusGraph on HBase and SQLG with H2 performance is in between.