As a special kind of graph database systems, RDF stores have been widely used in many applications, e.g., knowledge graphs and semantic web. RDF stores utilize SPARQL as their standardized query language to store and retrieve RDF graphs. Incorrect implementations of RDF stores can introduce logic bugs that return an incorrect query result for a given SPARQL query. These logic bugs can cause severe consequences and are likely to go unnoticed by developers. However, no available tools can detect logic bugs in RDF stores.

In this paper, we propose RD2, a Randomized Differential testing approach of RDF stores, to reveal discrepancies among RDF stores, which indicate potential logic bugs in RDF stores. The core idea of RD2 is to build an equivalent RDF graph for multiple RDF stores, and verify whether they can return the same query result for a given SPARQL query. To effectively test RDF stores, we combine the SPARQL syntax and the generated RDF graphs to construct syntactically valid SPARQL queries that can return non-empty query results with high probability. We further unify the formats of SPARQL query results for different RDF stores and find discrepancies among them. We evaluate RD2 on three popular and widely-used RDF stores. In total, we have found 7 unique discrepancies among them, and all of them have been confirmed be developers. A video demonstration of RD2 is available at https://youtu.be/bxA0PM9XrKg.