Meituan's Graph Database Selection and Platform Construction

Meituan faces massive graph data storage and multi‑hop query requirements, with billions of vertices and edges across various business scenarios such as knowledge graph mining, security risk control, link analysis, and organizational management.

The article outlines the selection process for a distributed, open‑source graph database, evaluating candidates against five criteria: open source, distributed architecture, millisecond‑level multi‑hop latency, support for trillions of edges, and bulk import capability. Three categories were identified: (1) single‑node solutions like Neo4j, (2) JanusGraph‑style systems that add a graph layer on existing stores, and (3) native distributed stores such as DGraph and NebulaGraph. Benchmarks on the LDBC‑SNB dataset showed NebulaGraph outperforming competitors in data import, real‑time writes, and multi‑hop queries.

NebulaGraph’s architecture consists of three services—Meta Service, Storage Service, and Query Service—implemented in C++. It adopts a shared‑nothing distributed storage layer, separates compute from storage, and uses Raft for strong consistency. The platform built on NebulaGraph provides four layers: data‑application, data‑storage, data‑production, and supporting services, offering high availability (AP mode), hourly billions‑scale data import via Spark‑generated SST files, real‑time multi‑cluster write synchronization through Kafka, and an interactive graph visualization component.

Four business use cases are presented: an intelligent assistant powered by a food‑service knowledge graph, a medical‑beauty search recall system, a graph‑based recommendation‑reason generation module, and a code‑dependency analysis tool that writes code‑level dependencies into the graph for impact analysis and testing.

In summary, the platform delivers a self‑service, highly available graph database solution that handles petabyte‑scale data with millisecond latency, supports massive data ingestion, ensures eventual consistency across clusters, and offers rich visualization and management features. Future work includes enhancing the core graph engine for stability and exploring graph learning and computation capabilities.