Unified Graph Query Engine for Heterogeneous Databases: Enabling High-Performance Graph Queries Without Data Migration

Abstract

Graph query paradigms, which are popularized by data engineering, are highly efficient with complex, connected data in comparison to traditional SQL. Multi-join operations are exponentially complex in relational systems. Graph benefits may need expensive migration or ETL to specialized databases, which are hard to use with heterogeneous environments. Based on PuppyGraph's zero-ETL querying over relational and lakehouse target sources, this paper introduces a single overall graph query engine as a lightweight overlay. This solution is compatible with a wide variety of backends (RDBMS, NoSQL (MongoDB, Cassandra), big data (Hadoop/Spark), columnar (ClickHouse), and parquet-based Iceberg tables). It eradicates migration, achieves better storage efficiency through compression, increases security through column-level encryption and snapshots, and increases performance. Significant innovations are conditional lazy materialization to optimize Parquet/Iceberg with high-hop (>5), long-running (>5 s), or repeated (>= 3 in 24 hours) queries; a federation layer to hybrid cross-database queries; and graph optimization, such as index-free adjacency and rewrite rules. Experiments on reduction in versions of TPC-H and LDBC SNB-like graphs indicate a 55-70% reduction in latency on 3-10+ hop traversals, and a 40% reduction in storage by Iceberg compression and selective materialization. This can be achieved through introducing graph analytics to polyglot infrastructures without interference, redundancy, and data movement.