Development

Here are some pointers for hacking on bezier.

• Development

  • Adding Features

  • libbezier

  • Binary Extension

    • Building / Compiling

    • Dependencies

  • Running Unit Tests

    • Testing the Binary Extension

    • Test Coverage

    • Slow Tests

    • Slow Install

  • Functional Tests

    • Functional Test Data

  • Coding Style

    • Docstring Style

  • Documentation

    • Documentation Snippets

    • Documentation Images

  • Continuous Integration

  • Release Process / Deploying New Versions

    • Supported Python Versions

    • Versioning

  • Environment Variables

Adding Features

In order to add a feature to bezier:

1. Discuss: File an issue to notify maintainer(s) of the proposed changes (i.e. just sending a large PR with a finished feature may catch maintainer(s) off guard).

2. Add tests: The feature must work fully on CPython versions 3.6, 3.7 and 3.8 and on PyPy 3; on Linux, macOS and Windows. In addition, the feature should have 100% line coverage.

3. Documentation: The feature must (should) be documented with helpful doctest examples wherever relevant.

libbezier

To build the Fortran shared library directly, use CMake version 3.1 or later:

$ SRC_DIR="src/fortran/"
$ BUILD_DIR=".../libbezier-debug/build"
$ INSTALL_PREFIX=".../libbezier-debug/usr"
$ mkdir -p "${BUILD_DIR}"
$ cmake \
>     -DCMAKE_BUILD_TYPE=Debug \
>     -DCMAKE_INSTALL_PREFIX:PATH="${INSTALL_PREFIX}" \
>     -DCMAKE_VERBOSE_MAKEFILE:BOOL=ON \
>     -S "${SRC_DIR}" \
>     -B "${BUILD_DIR}"
$ cmake \
>     --build "${BUILD_DIR}" \
>     --config Debug \
>     --target install

Binary Extension

Many low-level computations have alternate implementations in Fortran. See the Python Binary Extension page for a more detailed description.

Building / Compiling

To compile the binary extension (with libbezier built and installed already) run:

$ # One of
$ BEZIER_INSTALL_PREFIX=.../usr/ python -m pip wheel .
$ BEZIER_INSTALL_PREFIX=.../usr/ python -m pip install .
$ BEZIER_INSTALL_PREFIX=.../usr/ python setup.py build_ext
$ BEZIER_INSTALL_PREFIX=.../usr/ python setup.py build_ext --inplace

Using a Release build of libbezier may make debugging more difficult. Instead, a Debug build of libbezier will include debug symbols, without optimizations that move code around, etc.

To explicitly disable the building of the extension, the BEZIER_NO_EXTENSION environment variable can be used:

$ BEZIER_NO_EXTENSION=True .../bin/python -m pip wheel .

This environment variable is actually used for the nox -s docs session to emulate the RTD build environment (where no Fortran compiler is present).

Dependencies

Currently, the src/fortran/quadpack directory has a subset of Fortran 77 subroutines from QUADPACK. These are Public Domain, so they do not conflict with the Apache 2.0 license (as far as we know). In addition it contains another popular subroutine from NETLIB: d1mach (which the QUADPACK subroutines depend on).

QUADPACK is used to perform numerical quadrature to compute the length of a curve segment.

Running Unit Tests

We recommend using Nox to run unit tests:

$ nox -s "unit-3.6"
$ nox -s "unit-3.7"
$ nox -s "unit-3.8"
$ nox -s "unit-pypy3"
$ nox -s  unit  # Run all versions

However, pytest can be used directly (though it won’t manage dependencies or build the binary extension):

$ PYTHONPATH=src/python/ python3.6 -m pytest tests/unit/
$ PYTHONPATH=src/python/ python3.7 -m pytest tests/unit/
$ PYTHONPATH=src/python/ python3.8 -m pytest tests/unit/
$ PYTHONPATH=src/python/ pypy3     -m pytest tests/unit/

Testing the Binary Extension

When using nox, libbezier will be built and installed into a well-known BEZIER_INSTALL_PREFIX within the nox envdir (typically .nox/), the bezier package will automatically be installed into a virtual environment and the binary extension will be built during install.

However, if the tests are run directly from the source tree via

$ PYTHONPATH=src/python/ python -m pytest tests/unit/

some unit tests may be skipped. The unit tests that explicitly exercise the binary extension will skip (rather than fail) if the extension isn’t compiled (with build_ext --inplace) and present in the source tree.

Test Coverage

bezier has 100% line coverage. The coverage is checked on every build and uploaded to coveralls.io via the COVERALLS_REPO_TOKEN environment variable set in the CircleCI environment.

To run the coverage report locally:

$ nox -s cover
$ # OR
$ PYTHONPATH=src/python/ python -m pytest \
>     --cov=bezier \
>     --cov=tests.unit \
>     tests/unit/

Slow Tests

To run unit tests without test cases that have been (explicitly) marked slow, use the --ignore-slow flag:

$ nox -s "unit-3.6" -- --ignore-slow
$ nox -s "unit-3.7" -- --ignore-slow
$ nox -s "unit-3.8" -- --ignore-slow
$ nox -s  unit      -- --ignore-slow

These slow tests have been identified via:

$ ...
$ nox -s "unit-3.8" -- --durations=10

and then marked with pytest.mark.skipif.

Slow Install

Installing NumPy with PyPy can take upwards of two minutes and installing SciPy can take as much as seven minutes. This makes it prohibitive to create a new environment for testing.

In order to avoid this penalty, the WHEELHOUSE environment variable can be used to instruct nox to install NumPy and SciPy from locally built wheels when installing the pypy3 sessions.

To pre-build NumPy and SciPy wheels:

$ pypy3 -m virtualenv pypy3-venv
$ pypy3-venv/bin/python -m pip wheel --wheel-dir=${WHEELHOUSE} numpy
$ pypy3-venv/bin/python -m pip install ${WHEELHOUSE}/numpy*.whl
$ pypy3-venv/bin/python -m pip wheel --wheel-dir=${WHEELHOUSE} scipy
$ rm -fr pypy3-venv/

In addition to the WHEELHOUSE environment variable, the paths ${HOME}/wheelhouse and /wheelhouse will also be searched for pre-built wheels.

Alternatively, wheels can be downloaded from pypy-wheels, however the SciPy wheel will still require libatlas-dev, libblas-dev and liblapack-dev.

The Docker image for the CircleCI test environment has already pre-built these wheels and stored them in the /wheelhouse directory. So, in the CircleCI environment, the WHEELHOUSE environment variable is set to /wheelhouse.

Functional Tests

Line coverage and unit tests are not entirely sufficient to test numerical software. As a result, there is a fairly large collection of functional tests for bezier.

These give a broad sampling of curve-curve intersection, triangle-triangle intersection and segment-box intersection problems to check both the accuracy (i.e. detecting all intersections) and the precision of the detected intersections.

To run the functional tests:

$ nox -s "functional-3.6"
$ nox -s "functional-3.7"
$ nox -s "functional-3.8"
$ nox -s "functional-pypy3"
$ nox -s  functional  # Run all versions
$ # OR
$ PYTHONPATH=src/python/ python3.6 -m pytest tests/functional/
$ PYTHONPATH=src/python/ python3.7 -m pytest tests/functional/
$ PYTHONPATH=src/python/ python3.8 -m pytest tests/functional/
$ PYTHONPATH=src/python/ pypy3     -m pytest tests/functional/

For example, the following curve-curve intersection is a functional test case:

and there is a Curve-Curve Intersection document which captures many of the cases in the functional tests.

A triangle-triangle intersection functional test case:

a segment-box functional test case:

and a “locate point on triangle” functional test case:

Functional Test Data

The curve-curve and triangle-triangle intersection test cases are stored in JSON files:

• curves.json

• curve_intersections.json

• triangles.json

• triangle_intersections.json

This way, the test cases are programming language agnostic and can be repurposed. The JSON schema for these files are stored in the tests/functional/schema directory.

Coding Style

Code is PEP8 compliant and this is enforced with flake8 and Pylint.

To check compliance:

$ nox -s lint

A few extensions and overrides have been specified in the pylintrc configuration for bezier.

Docstring Style

We require docstrings on all public objects and enforce this with our lint checks. The docstrings mostly follow PEP257 and are written in the Google style, e.g.

Args:
    path (str): The path of the file to wrap
    field_storage (FileStorage): The :class:`FileStorage` instance to wrap
    temporary (bool): Whether or not to delete the file when the File
       instance is destructed

Returns:
    BufferedFileStorage: A buffered writable file descriptor

In order to support these in Sphinx, we use the Napoleon extension. In addition, the sphinx-docstring-typing Sphinx extension is used to allow for type annotation for arguments and result (introduced in Python 3.5).

Documentation

The documentation is built with Sphinx and automatically updated on RTD every time a commit is pushed to master.

To build the documentation locally:

$ nox -s docs
$ # OR (from a Python 3.6 or later environment)
$ PYTHONPATH=src/python/ ./scripts/build_docs.sh

Documentation Snippets

A large effort is made to provide useful snippets in documentation. To make sure these snippets are valid (and remain valid over time), doctest is used to check that the interpreter output in the snippets are valid.

To run the documentation tests:

$ nox -s doctest
$ # OR (from a Python 3.6 or later environment)
$ PYTHONPATH=src/python/:. sphinx-build -W \
>     -b doctest \
>     -d docs/build/doctrees \
>     docs \
>     docs/build/doctest

Documentation Images

Many images are included to illustrate the curves / triangles / etc. under consideration and to display the result of the operation being described. To keep these images up-to-date with the doctest snippets, the images are created as doctest cleanup.

In addition, the images in the Curve-Curve Intersection document and this document are generated as part of the functional tests.

To regenerate all the images:

$ nox -s docs_images
$ # OR (from a Python 3.6 or later environment)
$ export MATPLOTLIBRC=docs/ GENERATE_IMAGES=True PYTHONPATH=src/python/
$ sphinx-build -W \
>     -b doctest \
>     -d docs/build/doctrees \
>     docs \
>     docs/build/doctest
$ python tests/functional/make_segment_box_images.py
$ python tests/functional/make_triangle_locate_images.py
$ python tests/functional/make_curve_curve_images.py
$ python tests/functional/make_triangle_triangle_images.py
$ unset MATPLOTLIBRC GENERATE_IMAGES PYTHONPATH

Continuous Integration

Tests are run on CircleCI (Linux), Travis CI (macOS) and AppVeyor (Windows) after every commit. To see which tests are run, see the CircleCI config, the Travis config and the AppVeyor config.

On CircleCI, a Docker image is used to provide fine-grained control over the environment. There is a base python-multi Dockerfile that just has the Python versions we test in. The image used in our CircleCI builds (from bezier Dockerfile) installs dependencies needed for testing (such as nox and NumPy).

On Travis CI, Matthew Brett’s multibuild is used to install “official” python.org CPython binaries for macOS. Then tests are run in 64-bit mode (NumPy has discontinued 32-bit support).

On AppVeyor, the binary extension is built and tested with both 32-bit and 64-bit Python binaries.

Release Process / Deploying New Versions

New versions are pushed to PyPI manually after a git tag is created. The process is manual (rather than automated) for several reasons:

• The documentation and README (which acts as the landing page text on PyPI) will be updated with links scoped to the versioned tag (rather than master). This update occurs via the doc_template_release.py script.

• Several badges on the documentation landing page (index.rst) are irrelevant to a fixed version (such as the “latest” version of the package).

• The build badges in the README and the documentation will be changed to point to a fixed (and passing) build that has already completed (will be the build that occurred when the tag was pushed). If the builds pushed to PyPI automatically, a build would need to link to itself while being run.

• Wheels need be built for Linux, macOS and Windows. This process is becoming better, but is still scattered across many different build systems. Each wheel will be pushed directly to PyPI via twine.

• The release will be manually pushed to TestPyPI so the landing page can be visually inspected and the package can be installed from TestPyPI rather than from a local file.

Supported Python Versions

bezier explicitly supports:

• Python 3.6

• Python 3.7

• Python 3.8

• PyPy 3

Supported versions can be found in the noxfile.py config.

Versioning

bezier follows calendar versioning.

Environment Variables

This project uses environment variables for building the bezier._speedup binary extension:

• BEZIER_INSTALL_PREFIX: A directory where libbezier is installed, including the shared library (lib/) and headers (include/). This environment variable is required to build the binary extension.

• BEZIER_NO_EXTENSION: If set, this will indicate that only the pure Python package should be built and installed (i.e. without the binary extension).

• BEZIER_WHEEL: If set, this will indicate to setup.py that the current build is intended for a wheel. On Windows, this will involve renaming bezier.dll to a unique name (to avoid name collision) and updating _speedup*.pyd to refer to the new name.

for interacting with the system at import time:

• PATH: On Windows, we add the bezier/extra-dll package directory to the path so that the bezier.dll shared libary can be loaded at import time for Python versions before 3.8. After 3.8, modifying PATH no longer works for these purposes; the os.add_dll_directory() function is used.

and for running tests and interacting with Continuous Integration services:

• WHEELHOUSE: If set, this gives a path to prebuilt NumPy and SciPy wheels for PyPy 3.

• GENERATE_IMAGES: Indicates to nox -s doctest that images should be generated during cleanup of each test case.

• READTHEDOCS: Indicates currently running on Read The Docs (RTD). This is used to tell Sphinx to use the RTD theme when not running on RTD.

• COVERALLS_REPO_TOKEN: To upload the coverage report.