LDBC Social Network Benchmark

The LDBC Social Network Benchmark (SNB), from the Linked Data Benchmark Council, is the recognized standard for evaluating graph databases and graph query engines such as Neo4j, TigerGraph, Amazon Neptune, and others. Graph workloads — traversals, pattern matching, and path finding over highly connected data — are poorly captured by relational benchmarks, so LDBC built a benchmark around a realistic, correlated social network to fill that gap.

What It Measures

LDBC SNB has two main workloads. The Interactive workload measures transactional graph queries — short read traversals (such as a person's friends, recent posts, or shortest paths between people) mixed with updates — reporting query throughput and per-query latency including tail percentiles. The Business Intelligence workload measures complex analytical graph queries that touch large portions of the graph, reporting query times. There is also a Graphalytics component for graph algorithms like PageRank, BFS, and community detection.

Methodology

The benchmark uses a data generator that produces a synthetic but realistic social graph with correlated, skewed distributions — people, friendships, posts, comments, likes, and forums — at defined scale factors measured in gigabytes, scaling into terabytes. The generator deliberately introduces correlations (for example, people in the same city or with similar interests are more likely connected) so that query selectivity mirrors real graphs and cannot be gamed by uniform assumptions. A compliant Interactive run loads the data, then drives a query mix with both reads and updates at a target rate, measuring whether the system sustains the rate within latency bounds. The BI workload runs a set of heavy analytical queries. LDBC defines strict, auditable rules and provides reference implementations and drivers for fair comparison.

How to Interpret Results

State the scale factor and which workload (Interactive, BI, or Graphalytics) produced the numbers, as they measure very different things. For Interactive, read tail latency at the sustained throughput, since graph traversals can have long tails on highly connected hubs. For BI, compare per-query times across engines at the same scale. The data correlations mean LDBC rewards genuine graph-traversal efficiency and good index/storage design rather than relational tricks. Prefer audited results and verify the driver configuration and update mix.

Limitations

The benchmark models one domain — a social network — so it may not represent fraud-detection, knowledge-graph, or supply-chain graphs with different shapes. Full audited runs are operationally demanding, so many published numbers are unofficial subsets. Graph query languages differ across engines, complicating exact comparison. Use LDBC SNB to evaluate graph-database traversal and analytical performance on realistic, correlated data, complemented by tests on your own graph shape.