Building, Cython/C Extensions, and Releasing

The build process currently uses the Distribute package to build and install the astropy core (and any affiliated packages that use the template). The user doesn’t necessarily need to have distribute installed, as it will automatically bootstrap itself using the distribute_setup.py file in the source distribution if it isn’t installed for the user.

Customizing setup/build for subpackages

As is typical, there is a single setup.py file that is used for the whole astropy package. To customize setup parameters for a given sub-package, a setup_package.py file can be defined inside a package, and if it is present, the setup process will look for the following functions to customize the build process:

  • get_package_data()

    This function, if defined, should return a dictionary mapping the name of the subpackage(s) that need package data to a list of data file paths (possibly including wildcards) relative to the path of the package’s source code. e.g. if the source distribution has a needed data file astropy/wcs/tests/data/3d_cd.hdr, this function should return {'astropy.wcs.tests:'['data/3d_cd.hdr']}. See the package_data option of the distutils.core.setup() function.

    It is recommended that all such data be in a directory named “data” inside the package within which it is supposed to be used, and package data should be accessed via the astropy.utils.data.get_data_filename and astropy.utils.data.get_data_fileobj functions.

  • get_extensions()

    This provides information for building C or Cython extensions. If defined, it should return a list of distutils.core.Extension objects controlling the Cython/C build process (see below for more detail).

  • get_legacy_alias()

    This function allows for the creation of shims that allow a subpackage to be imported under another name. For example, astropy.io.fits used to be available under the namespace pyfits. For backward compatibility, it is helpful to have it still importable under the old name. Under most circumstances, this function should call astropy.setup_helpers.add_legacy_alias to generate a legacy module and then return what it returns.

  • get_build_options()

    This function allows a package to add extra build options. It should return a list of tuples, where each element has:

    • name: The name of the option as it would appear on the commandline or in the setup.cfg file.
    • doc: A short doc string for the option, displayed by setup.py build --help.
    • is_bool (optional): When True, the option is a boolean option and doesn’t have an associated value.

    Once an option has been added, its value can be looked up using astropy.setup_helpers.get_distutils_build_option.

  • get_external_libraries()

    This function declares that the package uses libraries that are included in the astropy distribution that may also be distributed elsewhere on the users system. It should return a list of library names. For each library, a new build option is created, --use-system-X which allows the user to request to use the system’s copy of the library. The package would typically call astropy.setup_helpers.use_system_library from its get_extensions function to determine if the package should use the system library or the included one.

The astropy.setup_helpers modules includes a update_package_files() function which automatically searches the given source path for setup_package.py modules and calls each of the above functions, if they exist. This makes it easy for affiliated packages to use this machinery in their own setup.py.

C or Cython Extensions

Astropy supports using C extensions for wrapping C libraries and Cython for speeding up computationally-intensive calculations. Both Cython and C extension building can be customized using the get_extensions() function of the setup_package.py file. If defined, this function must return a list of distutils.core.Extension objects. The creation process is left to the subpackage designer, and can be customized however is relevant for the extensions in the subpackage.

While C extensions must always be defined through the get_extensions() mechanism, Cython files (ending in .pyx) are automatically located and loaded in separate extensions if they are not in get_extensions(). For Cython extensions located in this way, headers for numpy C functions are included in the build, but no other external headers are included. .pyx files present in the extensions returned by get_extensions() are not included in the list of extensions automatically generated extensions. Note that this allows disabling a Cython file by providing an extension that includes the Cython file, but giving it the special name ‘cython_skip’. Any extension with this package name will not be built by setup.py.

Note

If an Extension object is provided for Cython source files using the get_extensions() mechanism, it is very important that the .pyx files be given as the source, rather than the .c files generated by Cython.

Installing C header files

If your C extension needs to be linked from other third-party C code, you probably want to install its header files along side the Python module.

  1. Create an include directory inside of your package for all of the header files.

  2. Use the get_package_data() hook in setup_package.py to install those header files. For example, the astropy.wcs package has this:

    def get_package_data():
        return {'astropy.wcs': ['include/*.h']}
    

Preventing importing at build time

In rare cases, some packages may need to be imported at build time. Unfortunately, anything that requires a C or Cython extension or processing through 2to3 will fail to import until the build phase has completed. In those cases, the _ASTROPY_SETUP_ variable can be used to determine if the package is being imported as part of the build and choose to not import problematic modules. _ASTROPY_SETUP_ is inserted into the builtins, and is True when inside of astropy’s setup.py script, and False otherwise.

For example, suppose there is a subpackage foo that needs to import a module called version.py at build time in order to set some version information, and also has a C extension, process, that will not be available in the source tree. In this case, astropy/foo/__init__.py would probably want to check the value of _ASTROPY_SETUP_ before importing the C extension:

if not _ASTROPY_SETUP_:
    from . import process

from . import version

Release

The current release procedure for Astropy involves a combination of an automated release script and some manual steps. Future versions will automate more of the process, if not all.

One of the main steps in performing a release is to create a tag in the git repository representing the exact state of the repository that represents the version being released. For Astropy we will always use signed tags: A signed tag is annotated with the name and e-mail address of the signer, a date and time, and a checksum of the code in the tag. This information is then signed with a GPG private key and stored in the repository.

Using a signed tag ensures the integrity of the contents of that tag for the future. On a distributed VCS like git, anyone can create a tag of Astropy called “0.1” in their repository–and where it’s easy to monkey around even after the tag has been created. But only one “0.1” will be signed by one of the Astropy project coordinators and will be verifiable with their public key.

Creating a GPG Signing Key and a Signed Tag

Git uses GPG to created signed tags, so in order to perform an Astropy release you will need GPG installed and will have to generated a signing key pair. Most *NIX installations come with GPG installed by default (as it is used to verify the integrity of system packages). If you don’t have the gpg command, consult the documentation for your system on how to install it.

For OSX, GPG can be installed from MacPorts using sudo port install gnupg.

To create a new public/private key pair, simply run:

$ gpg --gen-key

This will take you through a few interactive steps. For the encryption and expiry settings, it should be safe to use the default settings (I use a key size of 4096 just because what does a couple extra kilobytes hurt?) Enter your full name, preferably including your middle name or middle initial, and an e-mail address that you expect to be active for a decent amount of time. Note that this name and e-mail address must match the info you provide as your git configuration, so you should either choose the same name/e-mail address when you create your key, or update your git configuration to match the key info. Finally, choose a very good pass phrase that won’t be easily subject to brute force attacks.

If you expect to use the same key for some time, it’s good to make a backup of both your public and private key:

$ gpg --export --armor > public.key
$ gpg --export-secret-key --armor > private.key

Back up these files to a trusted location–preferably a write-once physical medium that can be stored safely somewhere. One may also back up their keys to a trusted online encrypted storage, though some might not find that secure enough–it’s up to you and what you’re comfortable with.

Add your public key to a keyserver

Now that you have a public key, you can publish this anywhere you like–in your e-mail, in a public code repository, etc. You can also upload it to a dedicated public OpenPGP keyserver. This will store the public key indefinitely (until you manually revoke it), and will be automatically synced with other keyservers around the world. That makes it easy to retrieve your public key using the gpg command-line tool.

To do this you will need your public key’s keyname. To find this enter:

$ gpg --list-keys

This will output something like:

/path/to/.gnupg/pubring.gpg
---------------------------------------------
pub   4096D/1234ABCD 2012-01-01
uid                  Your Name <your_email>
sub   4096g/567890EF 2012-01-01

The 8 digit hex number on the line starting with “pub”–in this example the “1234ABCD” unique keyname for your public key. To push it to a keyserver enter:

$ gpg --send-keys 1234ABCD

But replace the 1234ABCD with the keyname for your public key. Most systems come configured with a sensible default keyserver, so you shouldn’t have to specify any more than that.

Create a tag

Now test creating a signed tag in git. It’s safe to experiment with this–you can always delete the tag before pushing it to a remote repository:

$ git tag -s v0.1 -m "Astropy version 0.1"

This will ask for the password to unlock your private key in order to sign the tag with it. Confirm that the default signing key selected by git is the correct one (it will be if you only have one key).

Once the tag has been created, you can verify it with:

$ git tag -v v0.1

This should output something like:

object e8e3e3edc82b02f2088f4e974dbd2fe820c0d934
type commit
tag v0.1
tagger Your Name <your_email> 1339779534 -0400

Astropy version 0.1
gpg: Signature made Fri 15 Jun 2012 12:59:04 PM EDT using DSA key ID 0123ABCD
gpg: Good signature from "Your Name <your_email>"

You can use this to verify signed tags from any repository as long as you have the signer’s public key in your keyring. In this case you signed the tag yourself, so you already have your public key.

Note that if you are planning to do a release following the steps below, you will want to delete the tag you just created, because the release script does that for you. You can delete this tag by doing:

$ git tag -d v0.1

Release Procedure

The automated portion of the Astropy release procedure uses zest.releaser to create the tag and update the version. zest.releaser is extendable through hook functions–Astropy already includes a couple hook functions to modify the default behavior, but future releases may be further automated through the implementation of additional hook functions. In order to use the hooks, Astropy itself must be installed alongside zest.releaser. It is recommended to create a virtualenv specifically for this purpose.

This may seem like a lot of steps, but most of them won’t be necessary to repeat for each release. The advantage of using an automated or semi-automated procedure is that ensures a consistent release process each time.

  1. Update the list of contributors in the creditsandlicense.rst file. The easiest way to check this is do:

    $ git shortlog -s

    And just add anyone from that list who isn’t already credited.

  2. Install virtualenv if you don’t already have it. See the linked virtualenv documentation for details. Also, make sure that you have cython installed, as you will need it to generate the .c files needed for the release.

  3. Create and activate a virtualenv:

    $ virtualenv --system-site-packages --distribute astropy-release
    $ source astropy-release/bin/activate
  4. Obtain a clean version of the Astropy repository. That is, one where you don’t have any intermediate build files. Either use a fresh git clone or do git clean -dfx.

  5. Be sure you’re the “master” branch, and install Astropy into the virtualenv:

    $ python setup.py install

    This is necessary for two reasons. First, the entry points for the releaser scripts need to be availale, and these are in the Astropy package. Second, the build process will generate .c files from the Cython .pyx files, and the .c files are necessary for the source distribution.

  6. Install zest.releaser into the virtualenv; use --upgrade --force to ensure that the latest version is installed in the virtualenv (if you’re running a csh variant make sure to run rehash afterwards too):

    $ pip install zest.releaser --upgrade --force
  7. Ensure that all changes to the code have been committed, then start the release by running:

    $ fullrelease
  8. You will be asked to enter the version to be released. Press enter to accept the default (which will normally be correct) or enter a specific version string. A diff will then be shown of CHANGES.rst and setup.py showing that a release date has been added to the changelog, and that the version has been updated in setup.py. Enter ‘Y’ when asked to commit these changes.

  9. You will then be shown the command that will be run to tag the release. Enter ‘Y’ to confirm and run the command.

  10. When asked “Check out the tag (for tweaks or pypi/distutils server upload)” enter ‘N’: zest.releaser does not offer enough control yet over how the register and upload are performed so we will do this manually until the release scripts have been improved.

  11. You will be asked to enter a new development version. Normally the next logical version will be selected–press enter to accept the default, or enter a specific version string. Do not add ”.dev” to the version, as this will be appended automatically (ignore the message that says ”.dev0 will be appended”–it will actually be ”.dev” without the 0). For example, if the just-released version was “0.1” the default next version will be “0.2”. If we want the next version to be, say “1.0” then that must be entered manually.

  12. You will be shown a diff of CHANGES.rst showing that a new section has been added for the new development version, and showing that the version has been updated in setup.py. Enter ‘Y’ to commit these changes.

  13. When asked to push the changes to a remote repository, enter ‘Y’. This should complete the portion of the process that’s automated at this point.

  14. Check out the tag of the released version. For example:

    $ git checkout v0.1
  15. Create the source distribution by doing:

    $ python setup.py sdist

    Copy the produced .tar.gz somewhere and verify that you can unpack it, build it, and get all the tests to pass. It would be best to create a new virtualenv in which to do this.

  16. Register the release on PyPI with:

    $ python setup.py register
  17. Upload the source distribution to PyPI; this is preceded by re-running the sdist command, which is necessary for the upload command to know which distribution to upload:

    $ python setup.py sdist upload
  18. Update the website to reflect the fact there is now a stable release.

  19. Update Readthedocs so that it builds docs for the corresponding github tag, and set the default page to the new release.

  20. Create a bug fix branch. If the version just was released was a “X.Y.0” version (“0.1” or “0.2” for example–the final ”.0” is typically ommitted) it is good to create a bug fix branch as well. Starting from the tagged changset, just checkout a new branch and push it to the remote server. For example, after releasing version 0.1, do:

    $ git checkout -b v0.1.x

    Then edit setup.py so that the version is '0.1.1.dev', and commit that change. Then, do:

    $ git push upstream v0.1.x

Note

You may need to replace upstream here with astropy or whatever remote name you use for the main astropy repository.

The purpose of this branch is for creating bug fix releases like “0.1.1” and “0.1.2”, while allowing development of new features to continue in the master branch. Only changesets that fix bugs without making significant API changes should be merged to the bug fix branches.

  1. Create a bug fix label on GitHub; this should have the same name as the just created bug fix branch. This label should be applied to all issues that should be backported to the bug fix branch.

Creating a MacOS X Installer on a DMG

The bdist_dmg command can be used to create a .dmg disk image for MacOS X with a .pkg installer. In order to do this, you will need to ensure that you have the following dependencies installed:

To create a .dmg file, run:

python setup.py bdist_dmg

Note that for the actual release version, you should do this with the Python distribution from python.org (not e.g. MacPorts, EPD, etc.). The best way to ensure maximum compatibility is to make sure that Python and Numpy are installed into /Library/Frameworks/Python.framework using the latest stable .dmg installers available for those packages. In addition, the .dmg should be build on a MacOS 10.6 system, to ensure compatibility with 10.6, 10.7, and 10.8.

Before distributing, you should test out an installation of Python, Numpy, and Astropy from scratch using the .dmg installers, preferably on a clean virtual machine.

Future directions

We plan to switch to a newer packaging scheme when it’s more stable, the upcoming standard library packaging module, derived from the distutils2 project. Until it’s working right, however, we will be using distribute and distutils.