Pigweed’s Python build

Pigweed uses a custom GN-based build system to manage its Python code. The Pigweed Python build supports packaging, installation, and distribution of interdependent local Python packages. It also provides for fast, incremental static analysis and test running suitable for live use during development (e.g. with pw_watch) or in continuous integration.

Pigweed’s Python code is exclusively managed by GN, but the GN-based build may be used alongside CMake, Bazel, or any other build system. Pigweed’s environment setup uses GN to set up the initial Python environment, regardless of the final build system. As needed, non-GN projects can declare just their Python packages in GN.

Background

Developing software involves much more than writing source code. Software needs to be compiled, executed, tested, analyzed, packaged, and deployed. As projects grow beyond a few files, these tasks become impractical to manage manually. Build systems automate these auxiliary tasks of software development, making it possible to build larger, more complex systems quickly and robustly.

Python is an interpreted language, but it shares most build automation concerns with other languages. Pigweed uses Python extensively and must to address these needs for itself and its users.

Existing solutions

The Python programming langauge does not have an official build automation system. However, there are numerous Python-focused build automation tools with varying degrees of adoption. See the Python Wiki for examples.

A few Python tools have become defacto standards, including setuptools, wheel, and pip. These essential tools address key aspects of Python packaging and distribution, but are not intended for general build automation. Tools like PyBuilder and tox provide more general build automation for Python.

The Bazel build system has first class support for Python and other languages used by Pigweed, including protocol buffers.

Challenges

Pigweed’s use of Python is different from many other projects. Pigweed is a multi-language, modular project. It serves both as a library or middleware and as a development environment.

This section describes Python build automation challenges encountered by Pigweed.

Dependencies

Pigweed is organized into distinct modules. In Python, each module is a separate package, potentially with dependencies on other local or PyPI packages.

The basic Python packaging tools lack dependency tracking for local packages. For example, a package’s setup.py or setup.cfg lists all of its dependencies, but pip is not aware of local packages until they are installed. Packages must be installed with their dependencies taken into account, in topological sorted order.

To work around this, one could set up a private PyPI server instance, but this is too cumbersome for daily development and incompatible with editable package installation.

Testing

Tests are crucial to having a healthy, maintainable codebase. While they take some initial work to write, the time investment pays for itself many times over by contributing to the long-term resilience of a codebase. Despite their benefit, developers don’t always take the time to write tests. Any barriers to writing and running tests result in fewer tests and consequently more fragile, bug-prone codebases.

There are lots of great Python libraries for testing, such as unittest and pytest. These tools make it easy to write and execute individual Python tests, but are not well suited for managing suites of interdependent tests in a large project. Writing a test with these utilities does not automatically run them or keep running them as the codebase changes.

Static analysis

Various static analysis tools exist for Python. Two widely used, powerful tools are Pylint and Mypy. Using these tools improves code quality, as they catch bugs, encourage good design practices, and enforce a consistent coding style. As with testing, barriers to running static analysis tools cause many developers to skip them. Some developers may not even be aware of these tools.

Deploying static analysis tools to a codebase like Pigweed has some challenges. Mypy and Pylint are simple to run, but they are extremely slow. Ideally, these tools would be run constantly during development, but only on files that change. These tools do not have built-in support for incremental runs or dependency tracking.

Another challenge is configuration. Mypy and Pylint support using configuration files to select which checks to run and how to apply them. Both tools only support using a single configuration file for an entire run, which poses a challenge to modular middleware systems where different parts of a project may require different configurations.

Protocol buffers

Protocol buffers are an efficient system for serializing structured data. They are widely used by Google and other companies.

The protobuf compiler protoc generates Python modules from .proto files. protoc strictly generates protobuf modules according to their directory structure. This works well in a monorepo, but poses a challenge to a middleware system like Pigweed. Generating protobufs by path also makes integrating protobufs with existing packages awkward.

Requirements

Pigweed aims to provide high quality software components and a fast, effective, flexible development experience for its customers. Pigweed’s high-level goals and the challenges described above inform these requirements for the Pigweed Python build.

  • Integrate seamlessly with the other Pigweed build tools.

  • Easy to use independently, even if primarily using a different build system.

  • Support standard packaging and distribution with setuptools, wheel, and pip.

  • Correctly manage interdependent local Python packages.

  • Out-of-the-box support for writing and running tests.

  • Preconfigured, trivial-to-run static analysis integration for Pylint and Mypy.

  • Fast, dependency-aware incremental rebuilds and test execution, suitable for use with pw_watch.

  • Seamless protocol buffer support.

Detailed design

Build automation tool

Existing Python tools may be effective for Python codebases, but their utility is more limited in a multi-language project like Pigweed. The cost of bringing up and maintaining an additional build automation system for a single language is high.

Pigweed uses GN as its primary build system for all languages. While GN does not natively support Python, adding support is straightforward with GN templates.

GN has strong multi-toolchain and multi-language capabilities. In GN, it is straightforward to share targets and artifacts between different languages. For example, C++, Go, and Python targets can depend on the same protobuf declaration. When using GN for multiple languages, Ninja schedules build steps for all languages together, resulting in faster total build times.

Not all Pigweed users build with GN. Of Pigweed’s three supported build systems, GN is the fastest, lightest weight, and easiest to run. It also has simple, clean syntax. This makes it feasible to use GN only for Python while building primarily with a different system.

Given these considerations, GN is an ideal choice for Pigweed’s Python build.

Module structure

Pigweed Python code is structured into standard Python packages. This makes it simple to package and distribute Pigweed Python packages with common Python tools.

Like all Pigweed source code, Python packages are organized into Pigweed modules. A module’s Python package is nested under a py/ directory (see Module Structure).

Example layout of a Pigweed Python package.
module_name/
├── py/
│   ├── BUILD.gn
│   ├── setup.cfg
│   ├── setup.py
│   ├── pyproject.toml
│   ├── package_name/
│   │   ├── module_a.py
│   │   ├── module_b.py
│   │   ├── py.typed
│   │   └── nested_package/
│   │       ├── py.typed
│   │       └── module_c.py
│   ├── module_a_test.py
│   └── module_c_test.py
└── ...

The BUILD.gn declares this package in GN. For upstream Pigweed, a presubmit check in ensures that all Python files are listed in a BUILD.gn.

Pigweed prefers to define Python packages using setup.cfg files. In the above file tree setup.py and pyproject.toml files are stubs with the following content:

setup.py
import setuptools  # type: ignore
setuptools.setup()  # Package definition in setup.cfg
pyproject.toml
[build-system]
requires = ['setuptools', 'wheel']
build-backend = 'setuptools.build_meta'

The stub setup.py file is there to support running pip install --editable.

Each pyproject.toml file is required to specify which build system should be used for the given Python package. In Pigweed’s case it always specifies using setuptools.

See also

pw_python_package targets

The key abstraction in the Python build is the pw_python_package. A pw_python_package represents a Python package as a GN target. It is implemented with a GN template. The pw_python_package template is documented in Python GN templates.

The key attributes of a pw_python_package are

  • a setup.py file,

  • source files,

  • test files,

  • dependencies on other pw_python_package targets.

A pw_python_package target is composed of several GN subtargets. Each subtarget represents different functionality in the Python build.

  • <name> - Represents the Python files in the build, but does not take any actions. All subtargets depend on this target.

  • <name>.tests - Runs all tests for this package.

    • <name>.tests.<test_file> - Runs the specified test.

  • <name>.lint - Runs static analysis tools on the Python code. This is a group of two subtargets:

    • <name>.lint.mypy - Runs Mypy on all Python files, if enabled.

    • <name>.lint.pylint - Runs Pylint on all Python files, if enabled.

  • <name>.install - Installs the package in a Python virtual environment.

  • <name>.wheel - Builds a Python wheel for this package.

To avoid unnecessary duplication, all Python actions are executed in the default toolchain, even if they are referred to from other toolchains.

Testing

Tests for a Python package are listed in its pw_python_package target. Adding a new test is simple: write the test file and list it in its accompanying Python package. The build will run the it when the test, the package, or one of its dependencies is updated.

Static analysis

pw_python_package targets are preconfigured to run Pylint and Mypy on their source and test files. Users may specify which pylintrc and mypy.ini files to use on a per-package basis. The configuration files may also be provided in the directory structure; the tools will locate them using their standard means. Like tests, static analysis is only run when files or their dependencies change.

Packages may opt out of static analysis as necessary.

Installing packages in a virtual environment

Python packages declared in the Python build may be installed in a specified virtual environment. The default venv to use may be specified using a GN build arg. The venv may be overridden for individual targets. The Python build tracks installation status of packages based on which venv is in use.

The Python build examines the VIRTUAL_ENV environment variable to determine the current venv. If the selected virtual environment is active, packages are installed directly into it. If the venv is not active, it is activated before installing the packages.

Under construction

Pigweed Python targets are not yet aware of the virtual environment. Currently, all actions occur in the active venv without consulting VIRTUAL_ENV.

Python packages defined entirely in tree are installed with the --editable option. Partially or fully generated packages are installed without that option.

Building Python wheels

Wheels are the standard format for distributing Python packages. The Pigweed Python build supports creating wheels for individual packages and groups of packages. Building the .wheel subtarget creates a .whl file for the package using the PyPA’s build tool.

The .wheel subtarget of pw_python_package records the location of the generated wheel with GN metadata. Wheels for a Python package and its transitive dependencies can be collected from the pw_python_package_wheels key. See Python Distributables.

Protocol buffers

The Pigweed GN build supports protocol buffers with the pw_proto_library target (see pw_protobuf_compiler). Python protobuf modules are generated as standalone Python packages by default. Protocol buffers may also be nested within existing Python packages. In this case, the Python package in the source tree is incomplete; the final Python package, including protobufs, is generated in the output directory.

Generating setup.py

The pw_python_package target in the BUILD.gn duplicates much of the information in the setup.py or setup.cfg file. In many cases, it would be possible to generate a setup.py file rather than including it in the source tree. However, removing the setup.py would preclude using a direct, editable installation from the source tree.

Pigweed packages containing protobufs are generated in full or in part. These packages may use generated setup files, since they are always be packaged or installed from the build output directory.