Algeon 0.1.3

Algeon 0.1.3 adds native S3 Parquet read acceleration and read-only Apache Iceberg table support, fixes cuDF aggregation ordering, and updates the Flight SQL server container image.

Server image: ghcr.io/algeon-dev/algeon-server:0.1.3 carries the server with cuDF, cuGraph, cuVS, cuML, and Iceberg SQL for x86-64-v3 hosts, compiled for sm_80, sm_86, sm_90a, sm_100 and sm_120a, and needs an NVIDIA driver of R580 or newer. Mount a directory of Parquet files at /data; each Parquet file and each directory of Parquet files at its top level becomes a table at start. The Docker guide https://algeon.dev/build/docker covers running, configuring and smoke-testing the image, and building your own for other GPUs or features.

Features & Enhancements:

• Remote S3 Parquet reads: Native accelerated read path for remote S3 Parquet datasets.
• Apache Iceberg integration: Read-only Apache Iceberg table registration and query execution via algeon-iceberg.
• cuDF aggregation: Fixes ordering semantics for multi-column aggregation groups.

[dependencies]
algeon-datafusion = "=0.1.3"

All 14 crates are published to crates.io. Native RAPIDS libraries and Algeon shims are built from the v0.1.3 source tag https://github.com/algeon-dev/algeon/tree/v0.1.3. Release source: 73dadb70b. Image index digest: sha256:eda97085d7c4f05a40b4cbadebe4ad7061dd219763cee276e909e2fe8d19f242.

Algeon 0.1.2

Algeon 0.1.2 publishes the standalone Flight SQL server as a container image and adds Parquet table registration from a data directory.

Server image: ghcr.io/algeon-dev/algeon-server:0.1.2 carries the server with cuDF, cuGraph, cuVS and cuML SQL for x86-64-v3 hosts, compiled for sm_80, sm_86, sm_90a, sm_100 and sm_120a, and needs an NVIDIA driver of R580 or newer. Mount a directory of Parquet files at /data; each Parquet file and each directory of Parquet files at its top level becomes a table at start. The Docker guide covers running, configuring and smoke-testing the image, and building your own for other GPUs or features.

Server: the new [data] root setting registers the Parquet files and directories at the top level of a directory as tables before startup SQL runs; the image sets it to /data. Device enumeration errors now name a missing NVIDIA driver instead of reporting only a failed device count.

Build: build_libcudf.sh builds the curl that KvikIO links without Brotli, zstd, nghttp2, libidn2 or libssh2, so libcudf no longer picks up loader dependencies from whatever -dev packages the host has. build_libcugraph.sh fails before configuring when LD_LIBRARY_PATH would shadow the canonical libcuvs. The nightly toolchain the gate uses is pinned in .config/policies/nightly-toolchain.txt, and the build and prerequisites guides were corrected from a third-party build report.

[dependencies]
algeon-datafusion = "=0.1.2"

All 14 crates are published to crates.io. Native RAPIDS libraries and Algeon shims are built from the v0.1.2 source tag. Release source: 4e3fee0cc. Image index digest: sha256:900d87d1333664bacfbfc5951409d7ab93cede82db0a58ad844373ca1f5af88f.

Algeon 0.1.1

Algeon 0.1.1 is a documentation release. Each of the 14 crates now ships a README that states its contract for crates.io and docs.rs, the cuDF component design notes moved to components/cudf/DESIGN.md, and the public API snapshot of algeon-datafusion is checked in CI. Rustdoc for the cuML handle, allocation failure sources and random-forest stream pools was rewritten. There are no functional changes.

[dependencies]
algeon-datafusion = "=0.1.1"

Native RAPIDS libraries and Algeon shims are built from the matching v0.1.1 source tag. Release source: 2233c7e1d.

Algeon 0.1.0

Algeon 0.1.0 is the first crates.io release of the GPU query engine for Apache DataFusion. It supports cuDF relational execution and optional cuGraph, cuVS and cuML SQL functions. Algorithm results can feed joins, aggregations and other algorithms within a SQL statement.

Embed Algeon through the application entrypoint:

[dependencies]
algeon-datafusion = "=0.1.0"

The adapter has no default features. Enable cugraph, cuvs or cuml for the corresponding SQL functions, server for the embedded Flight SQL API, and nvml for optional device diagnostics. The standalone Flight SQL server is built from source.

The release contains 14 crates under the Apache-2.0 license. Native RAPIDS libraries and Algeon shims must be built separately from the matching v0.1.0 source tag. Applications require Linux, a supported NVIDIA GPU, a compatible CUDA toolkit and host compiler, and Rust 1.95 or newer. Repository source builds target x86-64-v3; external applications use their own Cargo CPU configuration.

Start with the application build guide, DataFusion session guide, or Flight SQL server guide. The API reference documents the published application API.

Each query attempt executes on one GPU. Input data, intermediate results and algorithm scratch space must fit its GPU memory budget. Parquet is the supported source format; persistent vector indexes and models are outside this release's scope.

Release source: b2eb2a705. Pre-publication validation covered all 14 archives, packaged rustdoc, 16 consumer build/link/GPU execution configurations and native discovery checks. GitHub CI passed the native-free checks and website build for this commit.

Post-publication checks downloaded all 14 crates and matched their SHA-256 hashes to the validated archives. An external application with a fresh lockfile resolved Algeon entirely from crates.io and passed build, loader and GPU execution checks with default features and with all release features. All 14 hosted docs.rs builds and versioned API pages passed verification.