Building distributions
Packaging your code into an sdist or a wheel.
A standard packaging format for Python code is the distribution: an archive that is published to a package index such as PyPI, and can be installed by pip. The two standard distribution archive types are sdists and wheels.
This page explains how to use Pants to build distributions from your code.
Typically, repositories without sophisticated tooling end up building a single distribution which includes the entire repo. But Pants makes it easy to create multiple distributions from the same repository.
Background: setuptools and PEP 517
For a long time, Setuptools was the de-facto standard mechanism for building Python distributions. Setuptools relies on a setup.py
script that you provide in your code. This script contains the instructions on what code to package into the distribution and what the requirements and other metadata of the distribution should be.
In the past few years, however, a new standard for specifying distribution builds has emerged: PEP 517. Under this standard (and its companion standard, PEP 518) you use pyproject.toml
to specify the python requirements and entry point for the builder code. This information is referred to as a build backend.
Examples of build backends include Setuptools, but also other systems with package-building capabilities, such as Flit or Poetry.
Pants reads a PEP 517 [build-system]
specification from pyproject.toml
and applies it to build your distributions. That is, Pants acts as a build frontend in PEP 517 parlance. It is common to continue to use Setuptools as the build backend, but doing so via PEP 517 lets you control the exact version of Setuptools that gets used, as well as any other requirements that must be present at build time.
If there is no pyproject.toml
with a [build-system]
table available, Pants falls back to using Setuptools directly.
The python_distribution
target
You configure a distribution using a python_distribution
target. This target provides Pants with the information needed to build the distribution.
PEP 517
If using a PEP 517 pyproject.toml
file, you might have a target layout similar to this:
resource(name="pyproject", source="pyproject.toml")
python_distribution(
name="mydist",
dependencies=[
":pyproject",
# Dependencies on code to be packaged into the distribution.
],
provides=python_artifact(
name="mydist",
version="2.21.0",
),
# Example of setuptools config, other build backends may have other config.
wheel_config_settings={"--global-option": ["--python-tag", "py37.py38.py39"]},
# Don't use setuptools with a generated setup.py.
# You can also turn this off globally in pants.toml:
#
# [setup-py-generation]
# generate_setup_default = false
generate_setup = False,
)
Running pants package example/dists:mydist
will cause Pants to inspect the [build-system]
table in pyproject.toml
, install the requirements specified in that table's requires
key, and then execute the entry point specified in the build-backend
key to build an sdist and a wheel, just as PEP 517 requires.
If you want to build just a wheel or just an sdist, you can set sdist=False
or wheel=False
on the python_distribution
target.
Setuptools
If relying on legacy Setuptools behavior, you don't have a pyproject.toml
resource, so your target is simply:
python_distribution(
name="mydist",
dependencies=[
# Dependencies on code to be packaged into the distribution.
],
provides=python_artifact(
name="mydist",
version="2.21.0",
),
wheel_config_settings={"--global-option": ["--python-tag", "py37.py38.py39"]},
)
Running pants package example/dists:mydist
will cause Pants to run Setuptools, which will in turn run the setup.py
script in the python_distribution
target's directory. If no such script exists, Pants can generate one for you (see below).
package
for other package formatsThis page focuses on building sdists and wheels with the pants package
goal. See package for information on other formats that can be built with pants package
, such as PEX binaries and zip/tar archives.
setup.py
Although alternatives exist, and PEP 517 enables them, Setuptools is still by far the most common choice for building distributions, whether via PEP 517 config, or directly via legacy support. If using Setuptools in either fashion, you need a setup.py
script alongside your python_distribution
target (and the target needs to depend on that script, typically via an explicit dependency on a python_sources
target that owns it).
You can either author setup.py
yourself (which is necessary if building native extensions), or have Pants generate one for you (see below).
By default, Pants will generate a setup.py
for every python_distribution
target, unless you set generate_setup = False
on the target. But you can flip this behavior by setting generate_setup_default = false
in the [setup-py-generation]
section of your pants.toml
config file. In that case Pants will only generate a setup.py
for python_distribution
targets that have generate_setup = True
set on them.
So if you expect to use handwritten setup.py
scripts for most distributions in your repo, you probably want to set generate-setup-default = false
and override it as needed. If you expect to mostly use generated setup.py
scripts, you can set generate-setup-default = true
(or just not set it, since that is the default).
setup.py
If you use a handwritten setup.py
, the generated distribution will have requirements on the packages you list in the install_requires
key, as expected. But Pants will not automatically use those as dependencies of the underlying sources, e.g., when running tests. They are strictly used when generating a distribution.
Instead, the "universe" of possible requirements of your source files must be specified as described here, and Pants will use dependency inference to select an appropriate subset as needed.
If Pants generates a setup.py
for you then the install_requires
value will be generated from the actual requirements of your source files.
Using a generated setup.py
Much of the data you would normally put in a setup.py
file is already known to Pants, so it can be convenient to let Pants generate setup.py
files for you, instead of maintaining them manually for each distributable project.
In this case, you may want to add some information to the provides=
field in the python_distribution
target, for Pants to place in the generated setup.py
:
python_distribution(
name="mydist",
dependencies=[
# Dependencies on code to be packaged into the distribution.
],
provides=python_artifact(
name="mydist",
version="2.21.0",
description="An example distribution built with Pants.",
author="Pantsbuild",
classifiers=[
"Programming Language :: Python :: 3.7",
],
),
wheel_config_settings={"--global-option": ["--python-tag", "py37.py38.py39"]},
)
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.
entry_points
field to register entry points like console_scripts
The entry_points
field allows you to configure setuptools-style entry points:
python_distribution(
name="my-dist",
entry_points={
"console_scripts": {"some-command": "project.app:main"},
"flake8_entry_point": {
"PB1": "my_flake8_plugin:Plugin",
"PB2": "my_flake8_plugin:AnotherPlugin",
},
provides=python_artifact(...),
)
Pants will infer dependencies on each entry point, which you can confirm by running pants dependencies path/to:python_dist
.
In addition to using the format path.to.module:func
, you can use an address to a pex_binary
target, like src/py/project:pex_binary
or :sibling_pex_binary
. Pants will use the entry_point
already specified by the pex_binary
, and it will infer a dependency on the pex_binary
target. This allows you to better DRY your project's entry points.
setup()
keyword argumentsYou 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
orversion
. - Run processes like Git to dynamically determine the
version
kwarg.
Start by reading about the Plugin API, then refer to the Custom python_artifact()
kwargs instructions.
Mapping source files to distributions
A Pants repo typically consists of one python_source
target per file (usually generated by several python_sources
targets). To build multiple distributions from the same repo, Pants must determine which libraries are bundled into each distribution.
In the extreme case, you could have one distribution per python_source
target, but publishing and consuming a distribution per file would of course not be practical. So in practice, multiple source files are bundled into a single distribution.
Naively, you might think that a python_distribution
publishes all the code of all the python_source
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.
If you use a handwritten setup.py
, you have to figure this out for yourself - Pants will bundle whatever the script tells it to. But if you let Pants generate setup.py
then it will apply the following algorithm:
Given a python_distribution
target D, take all the source files in the transitive dependency closure of D. Some of those source files 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_source
target S, the distribution in which S's code is published is chosen to be:
- A
python_distribution
that depends, directly or indirectly, on S. - Is S's closest filesystem ancestor among those satisfying 1.
If there are multiple such exported source files at the same degree of ancestry, the ownership
is ambiguous and an error is raised. If there is no python_distribution
that depends on S
and is its ancestor, then there is no owner and an error is raised.
This algorithm implies that all source files published by a distribution must be below it in the filesystem. It also guarantees that a source file is only published by a single distribution.
The generated setup.py
will have its install_requires
set to include the 3rdparty dependencies of the code bundled in the distribution, plus any other distributions from your own repo. For example, if distribution D1 contains code that has a dependency on some source file S, and that source file is published in distribution D2, then D1's requirements will include a dependency on D2. In other words, Pants does the right thing.
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_dependency_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_dependency_version_scheme = "compatible"
See https://www.python.org/dev/peps/pep-0440/#version-specifiers for more information on the ~=
specifier.
See publish for how to use Pants to publish distributions using Twine.