跳转至

关系核心图分析:以SQL规模查询图数据,以及为什么节点/边模型是性能税而非连接数据的真实写照

文章背景与核心概要

长期以来,业界普遍认为图分析必须依赖专用的图数据库引擎,而关系型数据库则根本不适合处理高度关联的数据。本文挑战了这一传统假设,指出对于现实企业的业务负载而言,采用图查询语言(如 Cypher)前端的列式关系型引擎不仅能媲美甚至超越原生图引擎,还能轻松突破内存型图系统常见的内存瓶颈。

作者并未将节点/边属性图视为连接数据的更优表象,而是认为它只是对关系表中已有关系的冗余重新编码。为此,作者推出了 ClickGraph 及其 Databricks 方言版本的同类工具 DeltaGraph。它们可将 Cypher 查询直接转换为原生关系型架构(表、列和外键),无需数据导入或隔离集群。由于其输出会转化为标准 SQL,性能优化依然十分直观。LDBC 社交网络基准测试套件(LDBC Social Network Benchmark)的测试结果以及与 Neo4j 等系统的对比(列式引擎性能高出数个数量级)有力地支持了这些观点。


摘要

Authors: Gene Zhang (PhD)
Submitted: September 1, 2026
Subjects: Databases (cs.DB), Artificial Intelligence (cs.AI), Programming Languages (cs.PL)
arXiv: 2609.01525 [cs.DB]
DOI: 10.48550/arXiv.2609.01525


Summary

This paper challenges the long-standing assumption that graph analytics requires dedicated, purpose-built graph engines and that relational databases are fundamentally unsuited for connected data. The author argues that for real-world enterprise workloads, columnar relational engines fronted by graph query languages (like Cypher) can match or exceed native graph engines while effortlessly scaling past the memory limits where in-memory graph systems typically fail.

Rather than viewing the node/edge property graph as a superior representation of connected data, the paper posits it as an unnecessary re-encoding of relationships already present in relational tables. To demonstrate this, the author introduces ClickGraph and its Databricks-dialect sibling DeltaGraph, which translate Cypher queries directly onto native relational schemas (tables, columns, and foreign keys) without requiring data imports or isolated clusters. Because the output translates to standard SQL, performance optimization remains accessible. Benchmarks across the LDBC Social Network Benchmark suite and comparisons against systems like Neo4j (where a columnar engine outperforms Neo4j by orders of magnitude) support these claims.

作者:Gene Zhang (博士) 提交时间:2026年9月01日 研究领域:数据库 (cs.DB)、人工智能 (cs.AI)、编程语言 (cs.PL) arXiv:2609.01525 [cs.DB] DOI:10.48550/arXiv.2609.01525


摘要

本文挑战了长期以来的一个假设:即图分析需要专用的、定制构建的图引擎,而关系型数据库根本不适合处理连接数据。作者认为,对于真实世界的企业负载而言,由图查询语言(如 Cypher)驱动的列式关系引擎可以匹配甚至超越原生图引擎,同时能够轻松突破内存图系统通常会遭遇失败的内存极限。

与其将节点/边属性图视为连接数据的更优表现形式,本文认为它只是对关系表中已经存在的关系进行的不必要的重新编码。为了证明这一点,作者介绍了 ClickGraph 及其 Databricks 方言的同类产品 DeltaGraph,它们将 Cypher 查询直接转换到原生关系模式(表、列和外键)上,而无需数据导入或隔离集群。由于输出会转换为标准 SQL,因此性能优化依然触手可及。LDBC 社交网络基准测试套件(LDBC Social Network Benchmark)的基准测试以及与 Neo4j 等系统的对比(其中列式引擎的性能比 Neo4j 高出几个数量级)有力地支持了这些观点。


核心亮点与论点

  • The Performance Tax of the Node/Edge Model: Property graphs re-encode relationship structures that often already exist explicitly via foreign keys in relational databases, introducing unnecessary query-time overhead.
  • Columnar Relational Scale: Relational engines bypass the memory walls of dedicated in-memory graph engines, handling enterprise-scale analytical graph queries effortlessly.
  • Zero-Import Architecture: Tools like ClickGraph and DeltaGraph execute Cypher queries directly in place on existing storage layers (such as ClickHouse, Databricks, or lakehouse files) with no separate cluster required.
  • Transparent Optimization: Because queries compile down to standard SQL, developers retain full visibility to optimize, rewrite, and extend execution paths.
  • Proven Performance: Backed by reproducible measurements across the LDBC Social Network Benchmark, demonstrating orders-of-magnitude performance advantages over traditional native graph engines like Neo4j.
  • 节点/边模型的性能税: 属性图重新编码了通常已经在关系数据库中通过外键显式存在的关系结构,从而引入了不必要的查询时开销。
  • 列式关系规模: 关系引擎绕过了专用内存图引擎的内存墙,能够轻松处理企业级分析图查询。
  • 零导入架构: 诸如 ClickGraphDeltaGraph 等工具可以直接在现有的存储层(如 ClickHouse、Databricks 或湖仓一体文件)上原地执行 Cypher 查询,而无需单独的集群。
  • 透明优化: 由于查询会被编译为标准 SQL,开发人员可以完全掌控执行路径,从而进行优化、重写和扩展。
  • 经过验证的性能: 以 LDBC 社交网络基准测试的可复现测量结果为后盾,展示出相比 Neo4j 等传统原生图引擎数个数量级的性能优势。