Obrain

A fast, embedded graph database engine in Rust — memory-mapped storage that is the on-disk format, crash-safe WAL, Cypher & GQL query languages, and native vector + full-text retrieval. Open a multi-million-node graph in milliseconds; run it in-process, as a CLI, or through language bindings.

Highlights

• Instant open — the .obrain format is directly memory-mappable: no deserialization, no rebuild. Measured cold opens: ~5 ms on a 1.35M-node / 1.9M-edge graph, ~17 ms on an 8.1M-node corpus.
• Crash-safe by construction — write-ahead log with atomic dual-header checkpoints; torn-write and SIGKILL recovery covered by an adversarial test suite (50-seed torn-WAL replay, idempotent re-application).
• O(1) analytics primitives — live label/edge-type counters and per-node typed-degree columns: count(n) in microseconds, top-K connected nodes without a scan.
• Query engines — openCypher and ISO GQL front-ends over a shared vectorized executor (chunked columns, LIMIT/label pushdowns).
• Retrieval built in — HNSW vector index, tiered quantized cascade (1-bit → PQ → f32, p95 ≤ 1 ms on 10⁶ nodes), BM25 full text.
• Bindings — Python, Node.js, WASM, Go, C, C#; plus a standalone CLI (obrain query/info/stats/backup/compact).
• Streaming importer — neo4j2obrain migrates a live Neo4j (bolt) into an obrain substrate.

Quickstart

# CLI
cargo install --path crates/obrain-cli
obrain init mygraph && obrain query mygraph "INSERT (:Person {name: 'Ada'})"
obrain query mygraph "MATCH (p:Person) RETURN p.name"

// Embedded
use obrain_engine::ObrainDB;

let db = ObrainDB::open("./mygraph")?;
let session = db.session();
session.execute("INSERT (:Person {name: 'Ada'})")?;
let rows = session.execute("MATCH (p:Person) RETURN p.name")?;

Architecture

 crates/
 ├── obrain-common      shared types (NodeId, Value, PropertyMap…)
 ├── obrain-core        graph traits, vectorized execution operators
 ├── obrain-substrate   mmap-native storage: records, zones, WAL, tiers
 ├── obrain-engine      ObrainDB, sessions, planners, Cypher/GQL translators
 ├── obrain-adapters    parsers, plugins, graph algorithms (Leiden, PageRank…)
 ├── obrain-reactive    post-commit mutation event bus
 ├── obrain-iam         identity & access (WAMI integration)
 ├── obrain-cli         command-line interface
 └── bindings/          python · node · wasm · go · c · csharp

The engine also carries the storage primitives (weighted edges, energy columns, activation counters) that power obrain-cognitive — a separate, closed layer that adds adaptive memory, learned retrieval and an LLM-facing runtime on top of this engine.

License & lineage

Apache-2.0. Obrain began as a fork of GrafeoDB/grafeo and has since been substantially rewritten (mmap-native substrate, WAL, tiered retrieval, GQL front-end); original copyright notices are preserved.