Pigweed’s GN Python Build#

How it Works#

In addition to compiler commands a Pigweed GN build will execute Python scripts for various reasons including running tests, linting code, generating protos and more. All these scripts are run as part of a pw_python_action GN template which will ultimately run python. Running Python on it’s own by default will make any Python packages installed on the users system available for importing. This is not good and can lead to flaky builds when different packages are installed on each developer workstation. To get around this the Python community uses virtual environments (venvs) that expose a specific set of Python packages separate from the host system.

When a Pigweed GN build starts a single venv is created for use by all pw_python_actions throughout the build graph. Once created, all required third-party Python packages needed for the project are installed. At that point no further modifications are made to the venv. Of course if a new third-party package dependency is added it will be installed too. Beyond that all venvs remain static. More venvs can be created with the pw_python_venv template if desired, but only one is used by default.

Every pw_python_action is run inside a venv

flowchart LR out[GN Build Dir<br/>fa:fa-folder out] out -->|ninja -C out| createvenvs createvenvs(Create venvs) createvenvs --> pyactions1 createvenvs --> pyactions2 subgraph pyactions1[Python venv 1] direction TB venv1(fa:fa-folder out/python-venv &nbsp) a1["pw_python_action('one')"] a2["pw_python_action('two')"] venv1 --> a1 venv1 --> a2 end subgraph pyactions2[Python venv 2] direction TB venv2(fa:fa-folder out/another-venv &nbsp) a3["pw_python_action('three')"] a4["pw_python_action('four')"] venv2 --> a3 venv2 --> a4 end

Note

Pigweed uses this venv target if a project does not specify it’s own build venv. See Build Time Python Virtualenv on how to define your own default venv.

Having a static venv containing only third-party dependencies opens the flood gates for python scripts to run. If the venv only contains third-party dependencies you may be wondering how you can import your own in-tree Python packages. Python code run in the build may still import any in-tree Python packages created with pw_python_package templates. However this only works if a correct python_deps arg is provided. Having that Python dependency defined in GN allows the pw_python_action to set PYTHONPATH so that given package can be imported. This has the benefit of the build failing if a dependency for any Python action or package is missing.

Benefits of Python venvs in GN

Using venvs to execute Python in GN provides reproducible builds with fixed third-party dependencies.
Using PYTHONPATH coupled with python_deps to import in-tree Python packages enforces dependency correctness.

Managing Python Requirements#

Build Time Python Virtualenv#

Pigweed’s GN Python build infrastructure relies on a single build-only venv for executing Python code. This provides an isolated environment with a reproducible set of third party Python constraints where all Python tests and linting can run. All pw_python_action targets are executed within this build venv.

The default build venv is specified via a GN arg and is best set in the root .gn or BUILD.gn file. For example:

pw_build_PYTHON_BUILD_VENV = "//:project_build_venv"

Third-party Python Requirements and Constraints#

Your project may have third party Python dependencies you wish to install into the bootstrapped environment and in the GN build venv. There are two main ways to add Python package dependencies:

Adding Requirements Files

Add a install_requires entry to a setup.cfg file defined in a pw_python_package template. This is the best option if your in-tree Python package requires an external Python package.
Create a standard Python requirements.txt file in your project and add it to the pw_build_PIP_REQUIREMENTS GN arg list.

Requirements files support a wide range of install locations including packages from pypi.org, the local file system and git repos. See pip’s Requirements File documentation for more info.

The GN arg can be set in your project’s root .gn or BUILD.gn file.
```
pw_build_PIP_REQUIREMENTS = [
  # Project specific requirements
  "//tools/requirements.txt",
]
```
See the GN Structure for Python Code section below for a full code listing.

Adding Constraints Files

Every project should ideally inherit Pigweed’s third party Python package version. This is accomplished via Python constraints files. Constraints control which versions of packages get installed by pip if that package is installed. To inherit Pigweed’s Python constraints include constraint.list from the pw_env_setup module from in your top level .gn file. Additonal project specific constraints can be appended to this list.

pw_build_PIP_CONSTRAINTS = [
  "$dir_pw_env_setup/py/pw_env_setup/virtualenv_setup/constraint.list",
  "//tools/constraints.txt",
]

GN Structure for Python Code#

Here is a full example of what is required to build Python packages using Pigweed’s GN build system. A brief file hierarchy is shown here with file content following. See also Pigweed Module Structure for Python Code below for details on the structure of Python packages.

Top level GN file hierarchy#

project_root/
├── .gn
├── BUILDCONFIG.gn
├── build_overrides/
│   └── pigweed.gni
├── BUILD.gn
│
├── python_package1/
│   ├── 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
│
├── third_party/
│   └── pigweed/
│
└── ...

project_root/

.gn

buildconfig = "//BUILDCONFIG.gn"
import("//build_overrides/pigweed.gni")

default_args = {
  pw_build_PIP_CONSTRAINTS = [
    # Inherit Pigweed Python constraints
    "$dir_pw_env_setup/py/pw_env_setup/virtualenv_setup/constraint.list",

    # Project specific constraints file
    "//tools/constraint.txt",
  ]

  pw_build_PIP_REQUIREMENTS = [
    # Project specific requirements
    "//tools/requirements.txt",
  ]

  # Default gn build virtualenv target.
  pw_build_PYTHON_BUILD_VENV = "//:project_build_venv"
}

BUILDCONFIG.gn

_pigweed_directory = {
  import("//build_overrides/pigweed.gni")
}

set_default_toolchain("${_pigweed_directory.dir_pw_toolchain}/default")

build_overrides / pigweed.gni

declare_args() {
  # Location of the Pigweed repository.
  dir_pigweed = "//third_party/pigweed/"
}

# Upstream Pigweed modules.
import("$dir_pigweed/modules.gni")

BUILD.gn

import("//build_overrides/pigweed.gni")

import("$dir_pw_build/python.gni")
import("$dir_pw_build/python_dist.gni")
import("$dir_pw_build/python_venv.gni")
import("$dir_pw_unit_test/test.gni")

# Lists all the targets build by default with e.g. `ninja -C out`.
group("default") {
  deps = [
    ":python.lint",
    ":python.tests",
  ]
}

# This group is built during bootstrap to setup the interactive Python
# environment.
pw_python_group("python") {
  python_deps = [
    # Generate and pip install _all_python_packages
    ":pip_install_project_tools",
  ]
}

# In-tree Python packages
_project_python_packages = [
  "//python_package1",
]

# Pigweed Python packages to include
_pigweed_python_packages = [
  "$dir_pw_env_setup:core_pigweed_python_packages",
  "$dir_pigweed/targets/lm3s6965evb_qemu/py",
  "$dir_pigweed/targets/stm32f429i_disc1/py",
]

_all_python_packages =
    _project_python_packages + _pigweed_python_packages

# The default venv for Python actions in GN
# Set this gn arg in a declare_args block in this file 'BUILD.gn' or in '.gn' to
# use this venv.
#
#   pw_build_PYTHON_BUILD_VENV = "//:project_build_venv"
#
pw_python_venv("project_build_venv") {
  path = "$root_build_dir/python-venv"
  constraints = pw_build_PIP_CONSTRAINTS
  requirements = pw_build_PIP_REQUIREMENTS

  # Ensure all third party Python dependencies are installed into this venv.
  # This works by checking the setup.cfg files for all packages listed here and
  # installing the packages listed in the [options].install_requires field.
  source_packages = _all_python_packages
}

# This template collects all python packages and their dependencies into a
# single super Python package for installation into the bootstrapped virtual
# environment.
pw_python_distribution("generate_project_python_distribution") {
  packages = _all_python_packages
  generate_setup_cfg = {
    name = "project-tools"
    version = "0.0.1"
    append_date_to_version = true
    include_default_pyproject_file = true
  }
}

# Install the project-tools super Python package into the bootstrapped
# Python venv.
pw_python_pip_install("pip_install_project_tools") {
  packages = [ ":generate_project_python_distribution" ]
}

Pigweed Module Structure for Python Code#

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 Pigweed Module Stucture).

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.

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 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.

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 any pw_python_package or pw_python_distribution 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 Distributable Templates.

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 packaged or installed from the build output directory.

Rationale#

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 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#

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.

Design Decision#

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.