Choosing between Python eclasses

Overview

The python-r1 eclass suite features 5 eclasses in total:

  1. python-utils-r1.eclass that provides utility functions common to all eclasses. This eclass is rarely inherited directly.

  2. python-any-r1.eclass that is suitable for packages using Python purely at build time.

  3. python-single-r1.eclass that provides a base for simpler packages installing Python scripts and alike.

  4. python-r1.eclass that provides a base for more complex packages, particularly these installing Python modules.

  5. distutils-r1.eclass that provides convenient phase functions and helpers for packages that primarily involve installing Python files.

_images/eclass.svg

Inheritance graph of python-r1 suite eclasses.

As a rule of thumb, the best eclass to use is the one that makes the ebuild the simplest while meeting its requirements. A more detailed process involves:

  1. Determining whether Python is used purely at build time, or at runtime as well. In the former case, python-any-r1 is the right choice.

  2. Determining whether single-impl or multi-impl approach is more appropriate. For the former, python-single-r1 is the correct base eclass. For the latter, python-r1.

  3. Determining whether the ebuild benefits from using distutils-r1. If it does, this eclass should be use instead (potentially along with DISTUTILS_SINGLE_IMPL to switch the underlying eclass).

Build time vs runtime use

The first basis for choosing Python eclass is whether Python is used merely at build time or at runtime as well.

A runtime use occurs if the package explicitly needs Python to be installed along with it, in order for it to function correctly. This generally happens if the package installs Python modules, extensions, scripts, or executables calling the Python interpreter or linking to libpython. This also applies to bash scripts or other executables that call python inline.

A build time use occurs if the package calls the Python interpreter or any kind of aforementioned executables during package’s build (or install) phases.

If the package uses Python purely at build time, the python-any-r1 eclass is appropriate. Otherwise, python-single-r1, python-r1 or distutils-r1 are to be used.

A specific exception to that rule is when the package is only calling external Python scripts directly (i.e. not via python /usr/bin/foo). If the called executables can be considered fully contained dependency-wise, there is no need to use an eclass.

For example, when using dev-util/meson to build a package, there is no need to use a Python eclass since Meson abstracts away its Pythonic implementation details and works as a regular executable for your packages. However, dev-util/scons requires Python eclass since it loads Python code from the package and a compatible Python version must be enforced.

Single-impl vs multi-impl

The second important basis for packages using Python at runtime is whether the package in question should support multi-implementation install or not.

A single-impl package is a package requiring the user to choose exactly one Python implementation to be built against. This means that the scripts installed by that package will be run via specified Python interpreter, and that the modules and extensions will be importable from it only. The package’s Python reverse dependencies will also have to use the same implementation. Since the package can’t support having more than one implementation enabled, its reverse dependencies have to be simple-impl as well.

Single-impl packages use python-single-r1 eclass. Writing ebuilds for them is easier since it is generally sufficient to call setup function early on, and the upstream build system generally takes care of using selected Python version correctly. Making packages single-impl is recommended when dealing with packages that are not purely written for Python or have single-impl dependencies.

A multi-impl package allows user to enable multiple (preferably any number of) implementations. The modules, extensions and scripts installed by the package are installed separately for each enabled implementation, and can therefore be used from any of them. The package can have reverse dependencies enabling only a subset of its implementations.

Multi-impl packages use python-r1 eclass. Ebuilds are more complex since they need to explicitly repeat build and install steps for each enabled implementation. Using this model is recommended for packages providing Python modules or extensions only, or having multi-impl reverse dependencies. In some cases supporting multi-impl build requires applying hacks, e.g. dev-libs/boost[python] uses non-standard names to install libboost_python for multiple Python versions.

The implementation for single-impl packages is selected via PYTHON_SINGLE_TARGET, while multi-impl uses PYTHON_TARGETS. These USE flag sets can be set independently to provide greater flexibility for developers and end users.

Both single-impl and multi-impl installs are supported by the distutils-r1 eclass.

Python-first packages (distutils-r1 eclass)

The third step in choosing the eclass for runtime use of Python is determining whether the ebuild would benefit from distutils-r1. This eclass is especially useful for packages that primarily focus on providing Python content. Its advantages include:

  • adding appropriate dependencies and REQUIRED_USE by default

  • a sub-phase function mechanism that makes installing Python modules in multi-impl mode easier

  • convenient support for building documentation using Sphinx and running tests using common Python test runners

In general, distutils-r1 should be preferred over the other eclasses if:

  • the package uses a PEP 517-compliant build system (i.e. has a pyproject.toml file with a build-system section)

  • the package uses a legacy distutils or setuptools build system (i.e. has a setup.py file)

  • the package primarily installs Python modules

In general, for multi-impl packages distutils-r1 is preferred over python-r1 as it usually makes the ebuilds simpler. For single-impl packages, python-single-r1 can sometimes be simpler.