Lynxes

A Fast, Zero-Copy Graph Analytics Engine Built Natively on Apache Arrow.

Why Lynxes | Quickstart | API Overview | Architecture

Lynxes is a blazingly fast, lazy-evaluated graph analytics engine. Unlike traditional Python libraries that wrap generic structures, Lynxes builds a graph-native engine directly over Arrow, completely bypassing the overhead of NetworkX or igraph.

Why Lynxes

• Zero-Copy Arrow Backing — NodeFrame and EdgeFrame directly own Apache Arrow RecordBatch. No intermediate copies, no Pandas/Polars dependency.
• Graph Structure as a First-Class Citizen — EdgeFrame always maintains a Compressed Sparse Row (CSR) index. Neighbor lookups are O(degree) from day one — no full table scans.
• Lazy by Default — No computation happens until you call .collect(). The built-in optimizer runs Predicate Pushdown, Projection Pushdown, Traversal Pruning, and Subgraph Caching before execution.
• Language-Agnostic Core — The query engine, storage engine, and graph algorithms are written entirely in Rust. Python is a thin zero-overhead PyO3 wrapper.

Quickstart

Install

pip install lynxes
# or
uv add lynxes

Build from source

git clone https://github.com/your-org/lynxes
cd lynxes/py-lynxes
uv run maturin develop --release

Python API

import lynxes as lx

# Load from .gf text, .gfb binary, or Parquet
g = lx.read_gf("graph.gf")
# g = lx.read_parquet_graph("nodes.parquet", "edges.parquet")
# g = lx.read_gfb("graph.gfb")

# Build a lazy plan — nothing executes yet
result = (
    g.lazy()
    .filter_nodes(lx.col("age") > 25)
    .expand("KNOWS", hops=2, direction="out")
    .aggregate_neighbors("KNOWS", lx.count().alias("friend_count"))
    .sort("friend_count", descending=True)
    .limit(10)
    .collect()
)

print(result)

Pattern Matching

Cypher-like pattern matching over the lazy execution engine:

result = (
    g.lazy()
    .match_pattern(
        [
            lx.node("person", "Person"),
            lx.edge("WORKS_AT"),
            lx.node("company", "Company"),
        ],
        where_=lx.col("person.age") > 25,
    )
    .collect()
)

Graph Algorithms

# PageRank
pr = g.pagerank()                          # → NodeFrame with 'pagerank' column

# Shortest path
path = g.shortest_path("alice", "charlie") # → ["alice", "bob", "charlie"]

# Connected components
cc = g.connected_components()              # → NodeFrame with 'component_id' column

# Betweenness centrality
bc = g.betweenness_centrality()

# Community detection (Louvain / Label Propagation)
cm = g.community_detection()

Remote Connectors

# Neo4j (Cypher)
g = lx.read_neo4j("bolt://localhost:7687", "neo4j", "password")

# ArangoDB (AQL)
g = lx.read_arangodb(
    endpoint="http://localhost:8529",
    database="mydb",
    graph="social",
    vertex_collection="persons",
    edge_collection="knows",
)

# SPARQL endpoint
g = lx.read_sparql(
    endpoint="https://dbpedia.org/sparql",
    node_template="SELECT ?id WHERE { ?id a <Thing> }",
    edge_template="SELECT ?s ?o WHERE { ?s ?p ?o }",
)

Distributed Graph Partitioning

# Partition a large graph across N shards
pg = g.partition(4, strategy="hash")   # or "range" / "label"
print(pg.n_shards)                     # 4
print(pg.stats())                      # imbalance ratio, boundary edges, …

# BFS across shard boundaries
nodes, edges = pg.distributed_expand(["alice"], hops=2, direction="out")

# Merge shards back into one GraphFrame
merged = pg.merge()

CLI

# Inspect a .gfb file
lynxes inspect graph.gfb

# Convert formats
lynxes convert graph.gf graph.gfb

# Run a filter query
lynxes query graph.gfb --filter "age > 25" --limit 10

API Overview

Top-level functions

GraphFrame methods

LazyGraphFrame methods

Architecture

Lynxes is organized as a multi-crate Rust workspace with a thin Python layer on top:

py-lynxes/                ← Python package (maturin / PyO3)
  src/lynxes/             ← lynxes Python namespace
  tests/unit/             ← pytest integration tests
  tests/benchmark/        ← NetworkX / igraph comparisons

crates/
  lynxes/                 ← Umbrella re-export crate
  lynxes-core/            ← Arrow frames, CSR index, algorithms,
  │                           expression types, logical plan, optimizer
  lynxes-plan/            ← Logical plan re-exports (thin)
  lynxes-io/              ← File I/O (.gf parser, .gfb binary, Parquet)
  lynxes-connect/         ← Remote connectors (Neo4j, ArangoDB,
  │                           SPARQL, Arrow Flight, GFConnector)
  lynxes-lazy/            ← LazyGraphFrame + query executor
  lynxes-python/          ← PyO3 binding crate (_lynxes.so)
  lynxes-cli/             ← `lynxes` command-line tool

Execution Pipeline

Python call
    │
    ▼
LazyGraphFrame (plan tree)
    │
    ▼
Optimizer ──── PredicatePushdown
            ── ProjectionPushdown
            ── TraversalPruning
            ── SubgraphCaching
            ── EarlyTermination
    │
    ▼
Executor ─────────────────────────────────────┐
    │                                         │
    ▼                                         ▼
NodeFrame / EdgeFrame                  CSR Index (O(degree))
(Arrow RecordBatch)                    BFS / Traversal / Algorithms

Crate Dependency Graph

lynxes-python ──┐
lynxes-cli    ──┤
                ├──► lynxes-lazy ──► lynxes-connect ──┐
                │                                      ├──► lynxes-io ──┐
                │                                      └──► lynxes-plan ─┤
                │                                                        ├──► lynxes-core
                └───────────────────────────────────────────────────────►┘

Documentation Map

• DESIGN.md — In-depth architectural design and engine principles
• docs/spec/ — Feature and restructure specifications
• py-lynxes/tests/benchmark/ — Performance benchmarks vs NetworkX / igraph

Contributing

Please read DESIGN.md first. Core principles that are non-negotiable:

1. Never wrap Polars — NodeFrame/EdgeFrame own Arrow RecordBatch directly
2. CSR is mandatory — EdgeFrame always holds a CSR index; no linear scan fallbacks
3. Lazy by default — All operations build a LogicalPlan; execution only on .collect()
4. No optimization without measurement — Run cargo bench before claiming speedups