pyproject RPM macros

These macros allow projects that follow the Python packaging specifications to be packaged as RPMs.

They work for:

• traditional Setuptools-based projects that use the setup.py file,
• newer Setuptools-based projects that have a setup.cfg file,
• general Python projects that use the PEP 517 pyproject.toml file (which allows using any build system, such as setuptools, flit or poetry).

These macros replace %py3_build and %py3_install, which only work with setup.py.

Usage

To use these macros, first BuildRequire the devel package for the Python you are building against. In Fedora, that's python3-devel.

BuildRequires: python3-devel

The macros will be pulled in as a dependency on Fedora and EPEL 9+.
In other distributions you need to BuildRequire the macros as well:

BuildRequires: python3-devel
BuildRequires: pyproject-rpm-macros

Next, you need to generate more build dependencies (of your projects and the macros themselves) by running %pyproject_buildrequires in the %generate_buildrequires section:

%generate_buildrequires
%pyproject_buildrequires

This will add build dependencies according to PEP 517 and PEP 518. This also adds run-time dependencies by default and can add test-time dependencies, see the section below. If you need more dependencies, such as non-Python libraries, BuildRequire them manually.

Note that %generate_buildrequires may produce error messages (exit 11) in the build log. This is expected behavior of BuildRequires generators; see the Fedora change for details.

Then, build a wheel in %build with %pyproject_wheel:

%build
%pyproject_wheel

And install the wheel in %install with %pyproject_install:

%install
%pyproject_install

%pyproject_install installs all wheels in pyproject-wheeldir/ located in the root of the source tree.

Adding run-time and test-time dependencies

To run tests or import checks in the %check section, the package's runtime dependencies need to also be included as build requirements.

Hence, %pyproject_buildrequires also generates runtime dependencies by default.

For this to work, the project's build system must support the prepare-metadata-for-build-wheel hook. The popular buildsystems (setuptools, flit, poetry) do support it.

This behavior can be disabled (e.g. when the project's build system does not support it) using the -R flag:

%generate_buildrequires
%pyproject_buildrequires -R

Alternatively, if the project specifies its dependencies in the pyproject.toml [project] table (as defined in PEP 621), the runtime dependencies can be obtained by reading that metadata.

This can be enabled by using the -p flag. This flag supports reading both the runtime dependencies, and the selected extras (see the -x flag described below).

Please note that not all build backends which use pyproject.toml support the [project] table scheme. For example, poetry-core (at least in 1.9.0) defines package metadata in the custom [tool.poetry] table which is not supported by the %pyproject_buildrequires macro.

Finally, the runtime dependencies can be obtained by building the wheel and reading the metadata from the built wheel. This can be enabled with the -w flag and cannot be combined with -p. Support for building wheels with %pyproject_buildrequires -w is provisional and the behavior might change. Please subscribe to Fedora's python-devel list if you use the option.

%generate_buildrequires
%pyproject_buildrequires -w

When this is used, the wheel is going to be built at least twice, becasue the %generate_buildrequires section runs repeatedly. To avoid accidentally reusing a wheel leaking from a previous (different) build, it cannot be reused between %generate_buildrequires rounds. Contrarily to that, rebuilding the wheel again in the %build section is redundant and the packager can omit the %build section entirely to reuse the wheel built from the last round of %generate_buildrequires. Be extra careful when attempting to modify the sources after %pyproject_buildrequires, e.g. when running extra commands in the %build section:

%build
cython src/wrong.pyx  # this is too late with %%pyproject_buildrequires -w
%pyproject_wheel

For projects that specify test requirements using an extra provide, these can be added using the -x flag. Multiple extras can be supplied by repeating the flag or as a comma separated list. For example, if upstream suggests installing test dependencies with pip install mypackage[testing], the test deps would be generated by:

%generate_buildrequires
%pyproject_buildrequires -x testing

For projects that specify test requirements using PEP 735 dependency groups, these can be added using the -g flag. Multiple groups can be supplied by repeating the flag or as a comma separated list. For example, if upstream uses a dependency group called tests, the test deps would be generated by:

%generate_buildrequires
%pyproject_buildrequires -g tests

For projects that specify test requirements in their tox configuration, these can be added using the -t flag (default tox environment) or the -e flag followed by the tox environment. The default tox environment (such as py37 assuming the Fedora's Python version is 3.7) is available in the %{toxenv} macro. For example, if upstream suggests running the tests on Python 3.7 with tox -e py37, the test deps would be generated by:

%generate_buildrequires
%pyproject_buildrequires -t

If upstream uses a custom derived environment, such as py37-unit, use:

%pyproject_buildrequires -e %{toxenv}-unit

Or specify more environments if needed:

%pyproject_buildrequires -e %{toxenv}-unit,%{toxenv}-integration

The -e option redefines %{toxenv} for further reuse. Use %{default_toxenv} to get the default value.

The -t/-e option uses tox-current-env's --print-deps-to-file behind the scenes. It generates dependencies listed directly in deps, dependencies defined through extras, and on tox 4.22+ also dependencies defined through dependency_groups.

If your package specifies some tox plugins in tox.requires, such plugins will be BuildRequired as well. Not all plugins are guaranteed to play well with tox-current-env, in worst case, patch/sed the requirement out from the tox configuration.

Note that neither -x or -t can be used with -R or -N, because runtime dependencies are always required for testing. You can only use those options if the build backend supports the prepare-metadata-for-build-wheel hook, or together with -p or -w. However, using -g with -R or -N is supported because dependency groups don't need to be used for testing and can be obtained by reading pyproject.toml only.

Additionally to generated requirements you can supply multiple file names to %pyproject_buildrequires macro. Dependencies will be loaded from them:

%pyproject_buildrequires requirements/tests.in requirements/docs.in requirements/dev.in

For packages not using build system you can use -N to entirely skip automatical generation of requirements and install requirements only from manually specified files. -N option implies -R and cannot be used in combination with other options mentioned above (-w, -e, -t, -x, -p).

The %pyproject_buildrequires macro also accepts the -r flag for backward compatibility; it means "include runtime dependencies" which has been the default since version 0-53.

Passing config settings to build backends

The %pyproject_buildrequires and %pyproject_wheel macros accept a -C flag to pass configuration settings to the build backend. Options take the form of -C KEY, -C KEY=VALUE, or -C--option-with-dashes. Pass -C multiple times to specify multiple options. This option is equivalent to pip's --config-settings flag. These are passed on to PEP 517 hooks' config_settings argument as a Python dictionary.

The %pyproject_buildrequires macro passes these options to the get_requires_for_build_wheel and prepare_metadata_for_build_wheel hooks. Passing -C to %pyproject_buildrequires is incompatible with -N which does not call these hooks at all.

The %pyproject_wheel macro passes these options to the build_wheel hook.

Consult the project's upstream documentation and/or the corresponding build backend's documentation for more information. Note that some projects don't use config settings at all and other projects may only accept config settings for one of the two steps.

Note that the current implementation of the macros uses pip to build wheels. On some systems (notably on RHEL 9 with Python 3.9), pip is too old to understand --config-settings. Using the -C option for %pyproject_wheel (or %pyproject_buildrequires -w) is not supported there and will result to an error like:

Usage:   
  /usr/bin/python3 -m pip wheel [options] <requirement specifier> ...
  ...
no such option: --config-settings

Running tox based tests

In case you want to run the tests as specified in tox configuration, you must use %pyproject_buildrequires with -t or -e as explained above. Then, use the %tox macro in %check:

%check
%tox

The macro:

• Sets environment variables via %{py3_test_envvars}, namely:
  • Always prepends $PATH with %{buildroot}%{_bindir}
  • If not defined, sets $PYTHONPATH to %{buildroot}%{python3_sitearch}:%{buildroot}%{python3_sitelib}
• If not defined, sets $TOX_TESTENV_PASSENV to *
• Runs tox with -q (quiet), --recreate and --current-env (from tox-current-env) flags
• Implicitly uses the tox environment name stored in %{toxenv} - as overridden by %pyproject_buildrequires -e

By using the -e flag, you can use a different tox environment(s):

%check
%tox
%if %{with integration_test}
%tox -e %{default_toxenv}-integration
%endif

If you wish to provide custom tox flags or arguments, add them after --:

%tox -- --flag-for-tox

If you wish to pass custom posargs to tox, use another --:

%tox -- --flag-for-tox -- --flag-for-posargs

Or (note the two sequential --s):

%tox -- -- --flag-for-posargs

Generating the %files section

To generate the list of files in the %files section, you can use %pyproject_save_files after the %pyproject_install macro. It takes toplevel module names (i.e. the names used with import in Python) and stores paths for those modules and metadata for the package (dist-info directory) to a file stored at %{pyproject_files}. For example, if a package provides the modules requests and _requests, write:

%install
%pyproject_install
%pyproject_save_files requests _requests

To add listed files to the %files section, use %files -f %{pyproject_files}. Note that you still need to add any documentation manually (for now).

%files -n python3-requests -f %{pyproject_files}
%doc README.rst

You can use globs in the module names if listing them explicitly would be too tedious:

%install
%pyproject_install
%pyproject_save_files '*requests'

In fully automated environments, you can use the * glob to include all modules (put it in single quotes to prevent Shell from expanding it). In Fedora however, you should always use a more specific glob to avoid accidentally packaging unwanted files (for example, a top level module named test).

Speaking about automated environments, some files cannot be classified with %pyproject_save_files, but it is possible to list all unclassified files by adding a special +auto argument.

%install
%pyproject_install
%pyproject_save_files '*' +auto

%files -n python3-requests -f %{pyproject_files}

However, in Fedora packages, always list executables explicitly to avoid unintended collisions with other packages or accidental missing executables:

%install
%pyproject_install
%pyproject_save_files requests _requests

%files -n python3-requests -f %{pyproject_files}
%doc README.rst
%{_bindir}/downloader

%pyproject_save_files can automatically mark license files with %license macro and language (*.mo) files with %lang macro and appropriate language code. Only license files declared via PEP 639 License-File field are detected. PEP 639 is still provisional and can be changed in the future. It is possible to use the -l flag to declare that a missing license should terminate the build or -L (the default) to explicitly disable this check. Packagers are encouraged to use the -l flag when the %license file is not manually listed in %files to avoid accidentally losing the file in a future version. When the %license file is manually listed in %files, packagers can use the -L flag to ensure future compatibility in case the -l behavior eventually becomes a default.

Note that %pyproject_save_files uses data from the RECORD file. If you wish to rename, remove or otherwise change the installed files of a package after %pyproject_install, %pyproject_save_files might break. If possible, remove/rename such files in %prep. If not possible, avoid using %pyproject_save_files or edit/replace %{pyproject_files}.

Performing an import check on all importable modules

If the upstream test suite cannot be used during the package build and you use %pyproject_save_files, you can benefit from the %pyproject_check_import macro. If %pyproject_save_files is not used, calling %pyproject_check_import will fail.

When %pyproject_save_files is invoked, it creates a list of all valid and public (i.e. not starting with _) importable module names found in the package. Each top-level module name matches at least one of the globs provided as an argument to %pyproject_save_files. This list is then usable by %pyproject_check_import which performs an import check for each listed module. When a module fails to import, the build fails.

The modules are imported from both installed and buildroot's %{python3_sitearch} and %{python3_sitelib}, not from the current directory.

Use the macro in %check:

%check
%pyproject_check_import

By using the -e flag, you can exclude module names matching the given glob(s) from the import check (put it in single quotes to prevent Shell from expanding it). The flag can be used repeatedly. For example, to exclude all submodules ending with config and all submodules starting with test, you can use:

%pyproject_check_import -e '*.config' -e '*.test*'

There must be at least one module left for the import check; if, as a result of greedy excluding, no modules are left to check, the check fails.

When the -t flag is used, only top-level modules are checked, qualified module names with a dot (.) are excluded. If the modules detected by %pyproject_save_files are requests, requests.models, and requests.packages, this will only perform an import of requests:

%pyproject_check_import -t

The modifying flags should only be used when there is a valid reason for not checking all available modules. The reason should be documented in a comment.

The %pyproject_check_import macro also accepts positional arguments with additional qualified module names to check, useful for example if some modules are installed manually. Note that filtering by -t/-e also applies to the positional arguments.

Another macro, %_pyproject_check_import_allow_no_modules allows to pass the import check, even if no Python modules are detected in the package. This may be a valid case for packages containing e.g. typing stubs. Don't use this macro in Fedora packages. It's only intended to be used in automated build environments such as Copr.

Generating Extras subpackages

The %pyproject_extras_subpkg macro generates simple subpackage(s) for Python extras.

The macro should be placed after the base package's %description to avoid issues in building the SRPM.

For example, if the requests project's metadata defines the extras security and socks, the following invocation will generate the subpackage python3-requests+security that provides python3dist(requests[security]), and a similar one for socks.

%pyproject_extras_subpkg -n python3-requests security socks

The macro works like %python_extras_subpkg, except the -i/-f/-F arguments are optional and discouraged. A filelist written by %pyproject_install is used by default. For more information on %python_extras_subpkg, see the Fedora change.

These arguments are still required:

• -n: name of the “base” package (e.g. python3-requests)
• Positional arguments: the extra name(s). Multiple subpackages are generated when multiple names are provided.

Provisional: Declarative Buildsystem (RPM 4.20+)

It is possible to reduce some of the spec boilerplate by using the provided pyproject declarative buildsystem. This option is only available with RPM 4.20+ (e.g. in Fedora 41+). The declarative buildsystem is provisional and the behavior might change. Please subscribe to Fedora's python-devel list if you use the feature.

To enable the pyproject declarative buildsystem, use the following:

BuildSystem:          pyproject
BuildOption(install): <options for %%pyproject_save_files>

That way, RPM will automatically fill-in the %prep, %generate_buildrequires, %build, %install, and %check sections the following defaults:

%prep
%autosetup -p1 -C

%generate_buildrequires
%pyproject_buildrequires

%build
%pyproject_wheel

%install
%pyproject_install
%pyproject_save_files <options from BuildOption(install)>

%check
%pyproject_check_import

To pass options to the individual macros, use BuildOption (see the documentation of declarative buildsystems).

# pass options for %%pyproject_save_files (mandatory when not overriding %%install)
BuildOption(install): -l _module +auto

# replace the default options for %%autosetup
BuildOption(prep): -S git_am -C

# pass options to %%pyproject_buildrequires
BuildOption(generate_buildrequires): docs-requirements.txt -t

# pass options to %%pyproject_wheel
BuildOption(build): -C--global-option=--no-cython-compile

# pass options to %%pyproject_check_import
BuildOption(check): -e '*.test*'

Alternatively, you can supply your own sections to override the automatic ones:

BuildOption(generate_buildrequires): -w
...
%build
# do nothing, the wheel was built in %%generate_buildrequires

You can append to end of the automatic sections:

%check -a
# run %%pytest after %%pyproject_check_import
%pytest

Or prepend to the beginning of them:

%prep -p
# run %%gpgverify before %%autosetup
%gpgverify -k2 -s1 -d0

Limitations

%pyproject_install changes shebang lines of every Python script in %{buildroot}%{_bindir} to #!%{__python3} %{py3_shbang_opt} (#!/usr/bin/python3 -s). Existing Python flags in shebangs are preserved. For example #!/usr/bin/python3 -Ru will be updated to #!/usr/bin/python3 -sRu. Sometimes, this can interfere with tests that run such scripts directly by name, because in tests we usually rely on PYTHONPATH (and -s ignores that). Would this behavior be undesired for any reason, undefine %{py3_shbang_opt} to turn it off.

Some valid Python version specifiers are not supported.

When a dependency is specified via an URL or local path, for example as:

https://github.com/ActiveState/appdirs/archive/8eacfa312d77aba28d483fbfb6f6fc54099622be.zip
/some/path/foo-1.2.3.tar.gz
git+https://github.com/sphinx-doc/sphinx.git@96dbe5e3

The %pyproject_buildrequires macro is unable to convert it to an appropriate RPM requirement and will fail. If the URL contains the packageName @ prefix as specified in PEP 508, the requirement will be generated without a version constraint:

appdirs@https://github.com/ActiveState/appdirs/archive/8eacfa312d77aba28d483fbfb6f6fc54099622be.zip
foo@file:///some/path/foo-1.2.3.tar.gz

Will be converted to:

python3dist(appdirs)
python3dist(foo)

Alternatively, when an URL requirement parsed from a text file given as positional argument to %pyproject_buildrequires contains the #egg=packageName fragment, as documented in pip's documentation:

git+https://github.com/sphinx-doc/sphinx.git@96dbe5e3#egg=sphinx

The requirements will be converted to package names without versions, e.g.:

python3dist(sphinx)

However upstreams usually only use direct URLs for their requirements as workarounds, so be prepared for problems.

Deprecated

The %{pyproject_build_lib} macro is deprecated, don't use it.

Testing the macros

This repository has two kinds of tests. First, there is RPM %check section, run when building the python-rpm-macros package.

Then there are CI tests. There is currently no way to run Fedora CI tests locally, but you can do what the tests do manually using mock. For each $PKG.spec in tests/:

• clean your mock environment:

  mock -r fedora-rawhide-x86_64 clean
  

• install the version of python-rpm-macros you're testing, e.g.:

  mock -r fedora-rawhide-x86_64 install .../python-rpm-macros-*.noarch.rpm
  

• download the sources:

  spectool -g -R $PKG.spec
  

• build a SRPM:

  rpmbuild -bs $PKG.spec
  

• build in mock, using the path from the command above as $SRPM:

  mock -r fedora-rawhide-x86_64 -n -N $SRPM