Algeon

Algeon is an experimental, open-source query engine from algeon-dev, built on Apache DataFusion. It exposes NVIDIA's cuGraph, cuVS and cuML libraries as SQL table functions and runs supported relational operations on cuDF. Analysts and AI agents can run graph, vector and machine-learning algorithms in a single SQL query. It requires Linux and a supported NVIDIA GPU.

What It Is

Algeon is a GPU-accelerated, single-process query engine for exploratory analysis of data in Parquet files. Each GPU algorithm is a table function whose input is a subquery, so the query that consumes a result also builds the graph or vectors. Results can be joined, filtered, aggregated or passed to the next function, with no separate graph or vector database and no export scripts. It can be used as a Rust library inside a DataFusion session or reached through an Arrow Flight SQL server with ADBC clients.

Algorithm Families

Three families of table functions wrap NVIDIA's GPU libraries:

• Graph analysis (cugraph_*): traversal, centrality, community detection, components, similarity, core structure and ForceAtlas2 layout.
• Vector analysis (cuvs_*): exact and approximate kNN, pairwise distances, per-row top-k, KMeans and PCA.
• Machine learning (cuml_*): fitting models within a query, or scoring features with a pretrained XGBoost, LightGBM or Treelite forest.

The project's MerRec demo composes personalized PageRank, brute-force kNN and a pretrained forest model in one query over Mercari marketplace events to select a campaign audience. Other demo datasets cover a citation graph, a product catalog and a malware benchmark.

How Execution Works

Queries enter through SQL, DataFusion DataFrames or Arrow Flight SQL. Algeon's DataFusion physical optimizer rule replaces supported plan fragments with in-process GPU execution on cuDF, and results come back as Arrow batches. Unsupported operators stay on DataFusion's CPU path. Execution modes include a default mode, functions_only (ordinary SQL stays on the CPU around explicit GPU functions) and native_required (rejects plans that still contain CPU operators). Explicit GPU functions have no CPU fallback, and failures return structured errors with retry guidance.

Scope and Current Status

The project is labeled experimental: APIs and query behavior may change incompatibly, and stability and result verification are incomplete, so outputs should be compared against trusted references. Each query runs on one GPU within a fixed memory grant with no spilling. GPU scans read Parquet from local disk or S3-compatible storage, read-only, and cover a subset of SQL and data types. Working sets larger than one GPU's memory, non-Parquet sources and long-lived indexes or models are documented as not a fit yet.

Setup Path

A published Docker server image serves Parquet files in a directory as tables over Flight SQL. Alternatively, build the Flight SQL server from source or embed Algeon in a Rust DataFusion session.