PythonPackage

Python packages and modules have their own special build system.

Phases

The PythonPackage base class provides the following phases that can be overridden:

  • build

  • build_py

  • build_ext

  • build_clib

  • build_scripts

  • clean

  • install

  • install_lib

  • install_headers

  • install_scripts

  • install_data

  • sdist

  • register

  • bdist

  • bdist_dumb

  • bdist_rpm

  • bdist_wininst

  • upload

  • check

These are all standard setup.py commands and can be found by running:

$ python setup.py --help-commands

By default, only the build and install phases are run:

  1. build - build everything needed to install

  2. install - install everything from build directory

If for whatever reason you need to run more phases, simply modify your phases list like so:

phases = ['build_ext', 'install', 'bdist']

Each phase provides a function <phase> that runs:

$ python -s setup.py --no-user-cfg <phase>

Each phase also has a <phase_args> function that can pass arguments to this call. All of these functions are empty except for the install_args function, which passes --prefix=/path/to/installation/prefix. There is also some additional logic specific to setuptools and eggs.

If you need to run a phase that is not a standard setup.py command, you’ll need to define a function for it like so:

phases = ['configure', 'build', 'install']

def configure(self, spec, prefix):
    self.setup_py('configure')

Important files

Python packages can be identified by the presence of a setup.py file. This file is used by package managers like pip to determine a package’s dependencies and the version of dependencies required, so if the setup.py file is not accurate, the package will not build properly. For this reason, the setup.py file should be fairly reliable. If the documentation and setup.py disagree on something, the setup.py file should be considered to be the truth. As dependencies are added or removed, the documentation is much more likely to become outdated than the setup.py.

Finding Python packages

The vast majority of Python packages are hosted on PyPI - The Python Package Index. pip only supports packages hosted on PyPI, making it the only option for developers who want a simple installation. Search for “PyPI <package-name>” to find the download page. Note that some pages are versioned, and the first result may not be the newest version. Click on the “Latest Version” button to the top right to see if a newer version is available. The download page is usually at: https://pypi.org/project/<package-name>

Description

The top of the PyPI downloads page contains a description of the package. The first line is usually a short description, while there may be a several line “Project Description” that follows. Choose whichever is more useful. You can also get these descriptions on the command-line using:

$ python setup.py --description
$ python setup.py --long-description

Homepage

Package developers use setup.py to upload new versions to PyPI. The setup method often passes metadata like homepage to PyPI. This metadata is displayed on the left side of the download page. Search for the text “Homepage” under “Project links” to find it. You should use this page instead of the PyPI page if they differ. You can also get the homepage on the command-line by running:

$ python setup.py --url

URL

You may have noticed that Spack allows you to add multiple versions of the same package without adding multiple versions of the download URL. It does this by guessing what the version string in the URL is and replacing this with the requested version. Obviously, if Spack cannot guess the version correctly, or if non-version-related things change in the URL, Spack cannot substitute the version properly.

Once upon a time, PyPI offered nice, simple download URLs like: https://pypi.python.org/packages/source/n/numpy/numpy-1.13.1.zip

As you can see, the version is 1.13.1. It probably isn’t hard to guess what URL to use to download version 1.12.0, and Spack was perfectly capable of performing this calculation.

However, PyPI switched to a new download URL format: https://pypi.python.org/packages/c0/3a/40967d9f5675fbb097ffec170f59c2ba19fc96373e73ad47c2cae9a30aed/numpy-1.13.1.zip#md5=2c3c0f4edf720c3a7b525dacc825b9ae

and more recently: https://files.pythonhosted.org/packages/b0/2b/497c2bb7c660b2606d4a96e2035e92554429e139c6c71cdff67af66b58d2/numpy-1.14.3.zip

As you can imagine, it is impossible for Spack to guess what URL to use to download version 1.12.0 given this URL. There is a solution, however. PyPI offers a new hidden interface for downloading Python packages that does not include a hash in the URL: https://pypi.io/packages/source/n/numpy/numpy-1.13.1.zip

This URL redirects to the files.pythonhosted.org URL. The general syntax for this pypi.io URL is: https://pypi.io/packages/source/<first-letter-of-name>/<name>/<name>-<version>.<extension>

Please use the pypi.io URL instead of the pypi.python.org URL. If both .tar.gz and .zip versions are available, .tar.gz is preferred. If some releases offer both .tar.gz and .zip versions, but some only offer .zip versions, use .zip.

PyPI vs. GitHub

Many packages are hosted on PyPI, but are developed on GitHub and other version control systems. The tarball can be downloaded from either location, but PyPI is preferred for the following reasons:

  1. PyPI contains the bare minimum of files to install the package.

    You may notice that the tarball you download from PyPI does not have the same checksum as the tarball you download from GitHub. When a developer uploads a new release to PyPI, it doesn’t contain every file in the repository, only the files necessary to install the package. PyPI tarballs are therefore smaller.

  2. PyPI is the official source for package managers like pip.

    Let’s be honest, pip is much more popular than Spack. If the GitHub tarball contains a file not present in the PyPI tarball that causes a bug, the developers may not realize this for quite some time. If the bug was in a file contained in the PyPI tarball, users would notice the bug much more quickly.

  3. GitHub release may be a beta version.

    When a developer releases a new version of a package on GitHub, it may not be intended for most users. Until that release also makes its way to PyPI, it should be assumed that the release is not yet ready for general use.

  4. The checksum for a GitHub release may change.

    Unfortunately, some developers have a habit of patching releases without incrementing the version number. This results in a change in tarball checksum. Package managers like Spack that use checksums to verify the integrity of a download tarball grind to a halt when the checksum for a known version changes. Most of the time, the change is intentional, and contains a needed bug fix. However, sometimes the change indicates a download source that has been compromised, and a tarball that contains a virus. If this happens, you must contact the developers to determine which is the case. PyPI is nice because it makes it physically impossible to re-release the same version of a package with a different checksum.

There are some reasons to prefer downloading from GitHub:

  1. The GitHub tarball may contain unit tests

    As previously mentioned, the PyPI tarball contains the bare minimum of files to install the package. Unless explicitly specified by the developers, it will not contain development files like unit tests. If you desire to run the unit tests during installation, you should use the GitHub tarball instead.

  2. Spack does not yet support spack versions and spack checksum with PyPI URLs

    These commands work just fine with GitHub URLs. This is a minor annoyance, not a reason to prefer GitHub over PyPI.

If you really want to run these unit tests, no one will stop you from submitting a PR for a new package that downloads from GitHub.

Build system dependencies

There are a few dependencies common to the PythonPackage build system.

Python

Obviously, every PythonPackage needs Python at build-time to run python setup.py build && python setup.py install. Python is also needed at run-time if you want to import the module. Due to backwards incompatible changes between Python 2 and 3, it is very important to specify which versions of Python are supported. If the documentation mentions that Python 3 is required, this can be specified as:

depends_on('python@3:', type=('build', 'run')

If Python 2 is required, this would look like:

depends_on('python@:2', type=('build', 'run')

If Python 2.7 is the only version that works, you can use:

depends_on('python@2.7:2.8', type=('build', 'run')

The documentation may not always specify supported Python versions. Another place to check is in the setup.py file. Look for a line containing python_requires. An example from py-numpy looks like:

python_requires='>=2.7,!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*'

More commonly, you will find a version check at the top of the file:

if sys.version_info[:2] < (2, 7) or (3, 0) <= sys.version_info[:2] < (3, 4):
    raise RuntimeError("Python version 2.7 or >= 3.4 required.")

This can be converted to Spack’s spec notation like so:

depends_on('python@2.7:2.8,3.4:', type=('build', 'run'))

setuptools

Originally, the Python language had a single build system called distutils, which is built into Python. Distutils provided a common framework for package authors to describe their project and how it should be built. However, distutils was not without limitations. Most notably, there was no way to list a project’s dependencies with distutils. Along came setuptools, a non-builtin build system designed to overcome the limitations of distutils. Both projects use a similar API, making the transition easy while adding much needed functionality. Today, setuptools is used in around 75% of the Python packages in Spack.

Since setuptools isn’t built-in to Python, you need to add it as a dependency. To determine whether or not a package uses setuptools, search the file for an import statement like:

import setuptools

or:

from setuptools import setup

Some packages are designed to work with both setuptools and distutils, so you may find something like:

try:
    from setuptools import setup
except ImportError:
    from distutils.core import setup

This uses setuptools if available, and falls back to distutils if not. In this case, you would still want to add a setuptools dependency, as it offers us more control over the installation.

Unless specified otherwise, setuptools is usually a build-only dependency. That is, it is needed to install the software, but is not needed at run-time. This can be specified as:

depends_on('py-setuptools', type='build')

cython

Compared to compiled languages, interpreted languages like Python can be quite a bit slower. To work around this, some Python developers rewrite computationally demanding sections of code in C, a process referred to as “cythonizing”. In order to build these package, you need to add a build dependency on cython:

depends_on('py-cython', type='build')

Look for references to “cython” in the setup.py to determine whether or not this is necessary. Cython may be optional, but even then you should list it as a required dependency. Spack is designed to compile software, and is meant for HPC facilities where speed is crucial. There is no reason why someone would not want an optimized version of a library instead of the pure-Python version.

Python dependencies

When you install a package with pip, it reads the setup.py file in order to determine the dependencies of the package. If the dependencies are not yet installed, pip downloads them and installs them for you. This may sound convenient, but Spack cannot rely on this behavior for two reasons:

  1. Spack needs to be able to install packages on air-gapped networks.

    If there is no internet connection, pip can’t download the package dependencies. By explicitly listing every dependency in the package.py, Spack knows what to download ahead of time.

  2. Duplicate installations of the same dependency may occur.

    Spack supports activation of Python extensions, which involves symlinking the package installation prefix to the Python installation prefix. If your package is missing a dependency, that dependency will be installed to the installation directory of the same package. If you try to activate the package + dependency, it may cause a problem if that package has already been activated.

For these reasons, you must always explicitly list all dependencies. Although the documentation may list the package’s dependencies, often the developers assume people will use pip and won’t have to worry about it. Always check the setup.py to find the true dependencies.

If the package relies on distutils, it may not explicitly list its dependencies. Check for statements like:

try:
    import numpy
except ImportError:
    raise ImportError("numpy must be installed prior to installation")

Obviously, this means that py-numpy is a dependency.

If the package uses setuptools, check for the following clues:

  • install_requires

    These packages are required for installation.

  • extra_requires

    These packages are optional dependencies that enable additional functionality. You should add a variant that optionally adds these dependencies.

  • test_requires

    These are packages that are required to run the unit tests for the package. These dependencies can be specified using the type='test' dependency type.

In the root directory of the package, you may notice a requirements.txt file. It may look like this file contains a list of all of the package’s dependencies. Don’t be fooled. This file is used by tools like Travis to install the pre-requisites for the package… and a whole bunch of other things. It often contains dependencies only needed for unit tests, like:

  • mock

  • nose

  • pytest

It can also contain dependencies for building the documentation, like sphinx. If you can’t find any information about the package’s dependencies, you can take a look in requirements.txt, but be sure not to add test or documentation dependencies.

setuptools

Setuptools is a bit of a special case. If a package requires setuptools at run-time, how do they express this? They could add it to install_requires, but setuptools is imported long before this and needed to read this line. And since you can’t install the package without setuptools, the developers assume that setuptools will already be there, so they never mention when it is required. We don’t want to add run-time dependencies if they aren’t needed, so you need to determine whether or not setuptools is needed. Grep the installation directory for any files containing a reference to setuptools or pkg_resources. Both modules come from py-setuptools. pkg_resources is particularly common in scripts in prefix/bin.

Passing arguments to setup.py

The default build and install phases should be sufficient to install most packages. However, you may want to pass additional flags to either phase.

You can view the available options for a particular phase with:

$ python setup.py <phase> --help

Each phase provides a <phase_args> function that can be used to pass arguments to that phase. For example, py-numpy adds:

def build_args(self, spec, prefix):
    args = []

    # From NumPy 1.10.0 on it's possible to do a parallel build.
    if self.version >= Version('1.10.0'):
        # But Parallel build in Python 3.5+ is broken.  See:
        # https://github.com/spack/spack/issues/7927
        # https://github.com/scipy/scipy/issues/7112
        if spec['python'].version < Version('3.5'):
            args = ['-j', str(make_jobs)]

    return args

Testing

PythonPackage provides a couple of options for testing packages.

Import tests

Just because a package successfully built does not mean that it built correctly. The most reliable test of whether or not the package was correctly installed is to attempt to import all of the modules that get installed. To get a list of modules, run the following command in the source directory:

$ python
>>> import setuptools
>>> setuptools.find_packages()
['numpy', 'numpy._build_utils', 'numpy.compat', 'numpy.core', 'numpy.distutils', 'numpy.doc', 'numpy.f2py', 'numpy.fft', 'numpy.lib', 'numpy.linalg', 'numpy.ma', 'numpy.matrixlib', 'numpy.polynomial', 'numpy.random', 'numpy.testing', 'numpy.core.code_generators', 'numpy.distutils.command', 'numpy.distutils.fcompiler']

Large, complex packages like numpy will return a long list of packages, while other packages like six will return an empty list. py-six installs a single six.py file. In Python packaging lingo, a “package” is a directory containing files like:

foo/__init__.py
foo/bar.py
foo/baz.py

whereas a “module” is a single Python file. Since find_packages only returns packages, you’ll have to determine the correct module names yourself. You can now add these packages and modules to the package like so:

import_modules = ['six']

When you run spack install --test=root py-six, Spack will attempt to import the six module after installation.

These tests most often catch missing dependencies and non-RPATHed libraries. Make sure not to add modules/packages containing the word “test”, as these likely won’t end up in installation directory.

Unit tests

The package you want to install may come with additional unit tests. By default, Spack runs:

$ python setup.py test

if it detects that the setup.py file supports a test phase. You can add additional build-time or install-time tests by overriding test and installtest, respectively. For example, py-numpy adds:

def install_test(self):
     with working_dir('..'):
         python('-c', 'import numpy; numpy.test("full", verbose=2)')

Setup file in a sub-directory

In order to be compatible with package managers like pip, the package is required to place its setup.py in the root of the tarball. However, not every Python package cares about pip or PyPI. If you are installing a package that is not hosted on PyPI, you may find that it places its setup.py in a sub-directory. To handle this, add the directory containing setup.py to the package like so:

build_directory = 'source'

Alternate names for setup.py

As previously mentioned, packages need to call their setup script setup.py in order to be compatible with package managers like pip. However, some packages like py-meep and py-adios come with multiple setup scripts, one for a serial build and another for a parallel build. You can override the default name to use like so:

def setup_file(self):
    return 'setup-mpi.py' if '+mpi' in self.spec else 'setup.py'

PythonPackage vs. packages that use Python

There are many packages that make use of Python, but packages that depend on Python are not necessarily PythonPackages.

Choosing a build system

First of all, you need to select a build system. spack create usually does this for you, but if for whatever reason you need to do this manually, choose PythonPackage if and only if the package contains a setup.py file.

Choosing a package name

Selecting the appropriate package name is a little more complicated than choosing the build system. By default, spack create will prepend py- to the beginning of the package name if it detects that the package uses the PythonPackage build system. However, there are occasionally packages that use PythonPackage that shouldn’t start with py-. For example:

  • busco

  • easybuild

  • httpie

  • mercurial

  • scons

  • snakemake

The thing these packages have in common is that they are command-line tools that just so happen to be written in Python. Someone who wants to install mercurial with Spack isn’t going to realize that it is written in Python, and they certainly aren’t going to assume the package is called py-mercurial. For this reason, we manually renamed the package to mercurial.

Likewise, there are occasionally packages that don’t use the PythonPackage build system but should still be prepended with py-. For example:

  • py-genders

  • py-py2cairo

  • py-pygobject

  • py-pygtk

  • py-pyqt

  • py-pyserial

  • py-sip

  • py-xpyb

These packages are primarily used as Python libraries, not as command-line tools. You may see C/C++ packages that have optional Python language-bindings, such as:

  • antlr

  • cantera

  • conduit

  • pagmo

  • vtk

Don’t prepend these kind of packages with py-. When in doubt, think about how this package will be used. Is it primarily a Python library that will be imported in other Python scripts? Or is it a command-line tool, or C/C++/Fortran program with optional Python modules? The former should be prepended with py-, while the latter should not.

extends vs. depends_on

This is very similar to the naming dilemma above, with a slight twist. As mentioned in the Packaging Guide, extends and depends_on are very similar, but extends adds the ability to activate the package. Activation involves symlinking everything in the installation prefix of the package to the installation prefix of Python. This allows the user to import a Python module without having to add that module to PYTHONPATH.

When deciding between extends and depends_on, the best rule of thumb is to check the installation prefix. If Python libraries are installed to prefix/lib/python2.7/site-packages (where 2.7 is the MAJOR.MINOR version of Python you used to install the package), then you should use extends. If Python libraries are installed elsewhere or the only files that get installed reside in prefix/bin, then don’t use extends, as symlinking the package wouldn’t be useful.

Alternatives to Spack

PyPI has hundreds of thousands of packages that are not yet in Spack, and pip may be a perfectly valid alternative to using Spack. The main advantage of Spack over pip is its ability to compile non-Python dependencies. It can also build cythonized versions of a package or link to an optimized BLAS/LAPACK library like MKL, resulting in calculations that run orders of magnitude faster. Spack does not offer a significant advantage to other python-management systems for installing and using tools like flake8 and sphinx. But if you need packages with non-Python dependencies like numpy and scipy, Spack will be very valuable to you.

Anaconda is another great alternative to Spack, and comes with its own conda package manager. Like Spack, Anaconda is capable of compiling non-Python dependencies. Anaconda contains many Python packages that are not yet in Spack, and Spack contains many Python packages that are not yet in Anaconda. The main advantage of Spack over Anaconda is its ability to choose a specific compiler and BLAS/LAPACK or MPI library. Spack also has better platform support for supercomputers. On the other hand, Anaconda offers Windows support.

External documentation

For more information on Python packaging, see: https://packaging.python.org/