Building distributions

Running setup.py commands to create an sdist or a wheel.

---

The standard Python packaging mechanism, setuptools, expects a build script called setup.py that describes how to package your project into a distribution: an archive that is uploaded to a package index such as PyPI, and can be installed by pip.

Common distribution formats include sdist, a source distribution format, and wheel, a built distribution format.

Much of the data you would normally put in a setup.py file is already known to Pants, so it's convenient to generate these files instead of maintaining them manually for each distributable project.

When run on appropriate targets, the ./pants package goal will generate a setup.py file, based on the information available in your BUILD files, and then run setup.py commands to package your code. For example, you can run the bdist_wheel command to create a wheel.

See here for more information about the setup.py file and the commands you can run with it.

See package for other package formats

This page focuses on building sdists and wheels with the ./pants package goal. See package for information on other formats that work with ./pants package, such as PEX binaries and zip/tar archives.

Benefit of Pants: multiple distributions in the same repository

Typically, repositories without sophisticated tooling end up with a single setup.py, which includes the entire repo in the distribution.

Pants makes it ergonomic to create multiple distributions from the same repository. Because Pants understands your code's true dependencies, each distribution will only be built with the exact code and metadata that it needs to work.

Configuring distributions

A distribution is set up using a python_distribution target, like this:

BUILD

python_distribution(
    name="mydist",
    dependencies=[
        ...
    ],
    provides=setup_py(
        name="mydist",
        version="2.21.0",
        description="An example distribution built with Pants.",
        author="Pantsbuild",
        classifiers=[
            "Programming Language :: Python :: 3.7",
        ],
    ),
    setup_py_commands=["sdist", "bdist_wheel", "--python-tag", "py36.py37"]
)

Some important setup.py metadata is inferred by Pants from your code and its dependencies. Other metadata needs to be provided explicitly. In Pants, as shown above, you do so through the provides field.

You can use almost any keyword argument accepted by setup.py in the setup() function.

However, you cannot use data_files, install_requires, namespace_packages, package_dir, package_data, or packages because Pants will generate these for you, based on the data derived from your code and dependencies.

Use .with_binaries() to register console_scripts entry points

You can use pex_binary targets to set the console_scripts value on your distribution.

The pex_binary target must explicitly set the field entry_point in the form path.to.module:my_func, e.g. project.app:main. See https://python-packaging.readthedocs.io/en/latest/command-line-scripts.html#the-console-scripts-entry-point.

To add a new entry point to your package, call the method .with_binaries(), like this:

provides=setup_py(
    ...
).with_binaries(my_command=":binary_target")

This will register a new console_script named my_command, which has the entry_point from the pex_binary target.

You may register multiple binary targets by using a dictionary, such as .with_binaries({"my_command": ":binary_target"}).

Consider writing a plugin to dynamically generate the setup() keyword arguments

You may want to write a plugin to do any of these things:

• Reduce boilerplate by hardcoding common arguments and commands.
• Read from the file system to dynamically determine kwargs, such as the long_description or version.
• Run processes like Git to dynamically determine the version kwarg.

Start by reading about the Plugin API, then refer to the Custom setup_py() kwargs instructions.

Mapping libraries to distributions

A Pants repo typically consists of many python_library targets in BUILD files spread across the codebase. To build distributions, Pants must determine which libraries are bundled into each distribution.

In the extreme case, you could have one distribution per library. But python_library targets tend to be fine-grained, representing perhaps a single Python package, whereas distributions are coarser grained, representing a larger project or piece of functionality. A codebase might have hundreds or thousands of python_library targets, but publishing and consuming that many distributions isn't practical.

So in practice, multiple libraries are typically bundled into a single distribution. Naively, you might think that a python_distribution publishes all the code of all the python_library targets it transitively depends on. But that could easily lead to trouble if you have multiple distributions that share common dependencies. You typically don't want the same code published in multiple distributions, as this can lead to all sorts of runtime import issues.

Instead, Pants applies the following algorithm:

Given a python_distribution target D, take all the libraries in the transitive dependency closure of D. Some of those libraries may be published in D itself, but others may be published in some other python_distribution target, D', in which case Pants will correctly add a requirement on D' in the metadata for D.

For each python_library target L, the distribution in which L's code is published is chosen to be:

1. A python_distribution that depends, directly or indirectly, on L.
2. Is L's closest filesystem ancestor among those satisfying 1.

If there are multiple such exported libraries at the same degree of ancestry, the ownership is ambiguous and an error is raised. If there is no python_distribution that depends on L and is its ancestor, then there is no owner and an error is raised.

This algorithm implies that all libraries published by a distribution must be below it in the filesystem. It also guarantees that a library is only published by a single distribution.

Changing the versioning scheme for first-party dependencies

When a python_distribution depends on another python_distribution, Pants will add it to the install_requires value in the generated setup.py.

By default, Pants will use exact requirements for first-party dependencies, like other_dist==1.0.1. You can set first_party_depenency_version_scheme in the [setup-py-generation] scope to 'compatible' to use ~= instead of ==, and any to leave off the version.

For example:

[setup-py-generation]
first_party_depenency_version_scheme = "compatible"

See https://www.python.org/dev/peps/pep-0440/#version-specifiers for more information on the ~= specifier.

Generating a setup.py file

If your python_distribution target has no setup_py_commands, then running ./pants package on that target will just create a directory with all of your package's code and a generated setup.py file.

The generated setup.py will have its install_requires set to include the 3rdparty dependencies of the code bundled in the distribution, and any other distributions from your own repo. For example, if distribution D1 contains code that has a dependency on some library L, and that library is published in distribution D2, then D1's requirements will include a dependency on D2. In other words, Pants does the right thing...

Running setup.py commands

If your python_distribution target has setup_py_commands, then running ./pants package on that target will run those commands for you.

For example, if setup_py_commands=["bdist_wheel", "sdist"], then ./pants package will build both a wheel and sdist for the given target.

How to upload your distributions to a package index

Pants does not have a mechanism to upload your distributions. Instead, run package as described above to create the asset, such as the wheel, then use a tool like Twine to upload your package.

Please let us know on Slack or GitHub issues if you would like this feature.