uv-bundler 1.0.0

Universal cross-platform Python packager powered by uv

uv-bundler

Build deployment-ready Python artifacts for Linux — from any machine. No Docker, no Linux build server.

Data pipelines (Spark, Flink, Lambda) run on Linux. Your laptop runs macOS. Getting the right platform wheels into a deployment artifact has always required a dedicated Linux build environment — until now.

---

The problem

• You need Linux wheels, but you're on macOS. A plain pip install fetches the wrong binaries for your host machine, not your target.
• Switching to ARM (AWS Graviton, Apple Silicon) is painful. Build pipelines are hard-coded to x86_64 and fail silently on aarch64.
• ZIP hacks break in production. Native extensions (NumPy, Pandas) crash with ImportError when bundled without the correct platform wheels.

---

The solution

uv-bundler uses Ghost Resolution — uv pip compile + uv pip install --python-platform — to resolve and download the correct Linux wheels on any host OS. No Docker, no cross-compilation, no remote build environment.

One command produces a self-contained artifact that runs on Linux x86_64 or aarch64:

uv-bundler --target spark-prod
# → dist/my-spark-job-linux-x86_64.jar

Run it on Linux with zero pre-installed packages:

python my-spark-job-linux-x86_64.jar
# Running single Spark job with pyspark 3.5.0

---

Quick start

1. Add a target to pyproject.toml:

[tool.uv-bundler]
project_name = "my-spark-job"
default_target = "spark-prod"

[tool.uv-bundler.targets.spark-prod]
format = "jar"
entry_point = "app.main:run"
platform = "linux"
arch = "x86_64"
python_version = "3.10"
manylinux = "2014"

2. Build from any machine:

# Verify config without building
uv-bundler --dry-run

# Build the artifact
uv-bundler --target spark-prod

3. Validate it works — no packages needed inside the container:

docker run --rm \
  -v "$(pwd)/dist:/artifacts" \
  python:3.10-slim \
  python /artifacts/my-spark-job-linux-x86_64.jar
# Running single Spark job with pyspark 3.5.0

---

Cross-arch builds

Build for a different architecture without any additional tooling:

# Build for ARM from any host (macOS, Linux x86_64, ...)
uv-bundler --target spark-prod --arch aarch64
# → dist/my-spark-job-linux-aarch64.jar

Ghost Resolution fetches manylinux2014_aarch64 wheels and bundles them directly — the host arch is irrelevant.

---

How it works

uv-bundler follows a 5-step lifecycle:

1. Context hydration — reads pyproject.toml, merges CLI overrides, validates the entry point module path.
2. Ghost Resolution — runs uv pip compile --python-platform <target> to pin platform-specific wheel hashes for the target OS, not the build host.
3. Staging — runs uv pip install --target to extract wheels into a staging directory alongside your source code.
4. Bootstrap generation — generates a __main__.py that correctly loads site-packages at runtime, including when executed directly from inside a zip archive (zipapp mode).
5. Assembly — bundles everything into the requested format.

---

Artifact formats

JAR — Spark / Flink

• Self-contained zipapp, runnable with python app.jar
• Includes META-INF/MANIFEST.MF for JVM tooling compatibility
• Dependencies bundled under site-packages/ inside the archive
• Build fails if .dylib or .dll binaries are detected (Linux targets only)

ZIP — AWS Lambda / general deployment

• Same internal layout as JAR (bootstrap + site-packages/ + sources)
• Suited for Lambda Layers and generic zip-based deployments

PEX — Airflow / schedulers

• Single-file executable Python environment via the pex CLI
• Uses a platform tag (e.g. manylinux2014_x86_64-cp-310-cp310) for cross-platform resolution
• First build downloads packages from PyPI; subsequent builds use the cache
• Cache $PEX_ROOT (default ~/.pex) in CI to avoid repeated downloads

---

Configuration reference

[tool.uv-bundler]
project_name = "analytics-engine"    # Used in artifact filename
default_target = "spark-prod"         # Target used when --target is omitted
output_base = "./dist"                # Output directory (override via OUT_DIR env var)

[tool.uv-bundler.targets.spark-prod]
# Core
format = "jar"                        # "jar" | "pex" | "zip"
entry_point = "app.main:run"          # module.submodule:function

# Resolution (cross-platform)
platform = "linux"                    # "linux" | "macos" | "windows"
arch = "x86_64"                       # "x86_64" | "aarch64"
python_version = "3.10"
manylinux = "2014"                    # "2010" | "2014" | numeric (e.g. "31")

# Packaging
compression = "deflated"              # "stored" | "deflated"

# Advanced
exclude = ["tests/*", "**/__pycache__", "*.pyc", ".git*"]
extra_files = { "config/prod.yaml" = "resources/config.yaml" }

OUT_DIR environment variable overrides output_base and takes precedence over the TOML value.

---

CLI reference

uv-bundler [OPTIONS]

Options:
  --target TEXT           Target name from [tool.uv-bundler.targets.*]
  --platform TEXT         Override platform  (linux | macos | windows)
  --arch TEXT             Override arch      (x86_64 | aarch64)
  --python-version TEXT   Override Python version (e.g. 3.10)
  --manylinux TEXT        Override manylinux tag  (e.g. 2014)
  --dry-run               Print resolved config without building

---

Installation

pip install uv-bundler
# or
uv tool install uv-bundler   # installs uv-bundler as an isolated global CLI

---

Installation (development)

git clone https://github.com/amarlearning/uv-bundler.git
cd uv-bundler

# Install uv — https://docs.astral.sh/uv/
brew install uv   # macOS

make setup        # creates .venv and installs all dev deps

make fmt          # format
make lint         # ruff + mypy
make test         # unit + integration tests

---

Contributing

• Branch: feature/<desc> or fix/<desc>
• Commit: imperative mood, ≤ 50 chars summary
• Quality gate: make fmt && make lint && make type && make test
• PR: include rationale; update docs if behavior changes

---

License

Apache License 2.0