Cypher graphs, from a unit test to bigger than RAM.

A graph database in Rust for developers who want Cypher inside their application: embed it as a library, or run the server and connect with the Neo4j drivers you already use.

Run it with the movies graph

$ docker run --rm -p 127.0.0.1:7475:7475 -p 127.0.0.1:7687:7687 ghcr.io/janvencl/graphbuff --demo movies

Needs Docker. Try it in your browser · Other ways to install

MATCH (k:Person {name: 'Keanu Reeves'})-[:ACTED_IN]->(m)<-[:ACTED_IN]-(co)
RETURN co.name, count(m) AS movies

co.namemovies

"Carrie-Anne Moss"2

"Hugo Weaving"1

What it is used for.

Each comes with a dataset to load and three queries to run.

Embedded in your program

A graph inside your Rust process: no server to start, no container in the test suite.

Recommendations

"Customers who bought this also bought" as a two-hop pattern, ranked in the same query.

Fraud and network analysis

Rings, shared devices and connected components, with the graph algorithms built in and a GPU backend for large graphs.

Knowledge graphs larger than RAM

Disk-resident storage with a page cache: the graph can outgrow memory and the queries stay the same.

21ms

Breadth-first search over 13 million relationships. On a gaming GPU.

Build with the gpu feature and GraphBuff moves its graph kernels onto any GPU with Vulkan, Direct3D 12 or Metal. No CUDA, no vendor lock-in, no separate analytics cluster. The planner decides per query, and the answers are the CPU's answers, down to the bit for PageRank.

24× faster BFS
9.5× faster multi-hop walk counts
5× faster Cypher count(*)
0 changes to your queries

Hardware, commit and command for every number

How the GPU backend works

Breadth-first search510 ms21 ms24×

Three-hop walk counts180 ms19 ms9.5×

Cypher two-hop count(*)154 ms31 ms5.0×

PageRank, 20 iterations1,293 ms557 ms2.3×

13 million relationships. Each row scaled to its own CPU time. Reproduce with cargo run --release -p graphbuff --features gpu --example gpubench.

Connect with the drivers you already use.

The server speaks Bolt 4.4 and 5.0 to 5.4, the protocol of the official Neo4j drivers. Smoke-tested with the Python and Java drivers. Prefer plain HTTP? Every query is one JSON POST.

Neo4j drivers
Python
Java
Rust

from neo4j import GraphDatabase

driver = GraphDatabase.driver(
    "bolt://127.0.0.1:7687", auth=("neo4j", "s3cret"))
records, summary, keys = driver.execute_query(
    "MATCH (p:Person) RETURN p.name AS name")

Three storage engines behind one API.

Choose per database when you open it. Queries, drivers and tools do not change.

Open source, with transactions, access control, TLS and backups in the box.

Open source: Free and open source under the GNU AGPL v3. Use it commercially. Project status
Snapshot reads: Readers never wait for writers. Each query sees one committed state, in every storage mode. Concurrency
Roles: reader, editor and admin, checked on every statement. PBKDF2 password hashes. Security
TLS 1.3: HTTPS and Bolt over TLS on the same ports, through rustls. Users and TLS
Online backup: Back up while queries keep committing. Restore and consistency checks from the CLI or over HTTP. Backup
Bulk import: 1 million relationships from CSV in about 9 seconds, in the neo4j-admin import format. Benchmark

All features

Two commands to a running graph.

$ docker run --rm -p 127.0.0.1:7475:7475 -p 127.0.0.1:7687:7687 ghcr.io/janvencl/graphbuff --demo movies

Needs Docker. Full walkthrough

Star on GitHub All downloads