Supporting multiple Python versions

Page Status:Incomplete
Last Reviewed:2014-12-24

Useful projects/resources to reference:

- DONE six
- DONE python-future (
- tox
- DONE Travis and Shining Panda CI (Shining Panda no longer available)
- DONE Appveyor
- DONE Ned Batchelder's "What's in Which Python"
- Lennart Regebro's "Porting to Python 3"
- Greg Hewgill's script to identify the minimum version of Python
  required to run a particular script:
- the Python 3 porting how to in the main docs
- cross reference to the stable ABI discussion
  in the binary extensions topic (once that exists)
- mention version classifiers for distribution metadata

In addition to the work required to create a Python package, it is often necessary that the package must be made available on different versions of Python. Different Python versions may contain different (or renamed) standard library packages, and the changes between Python versions 2.x and 3.x include changes in the language syntax.

Performed manually, all the testing required to ensure that the package works correctly on all the target Python versions (and OSs!) could be very time-consuming. Fortunately, several tools are available for dealing with this, and these will briefly be discussed here.

Automated Testing and Continuous Integration

Several hosted services for automated testing are available. These services will typically monitor your source code repository (e.g. at Github or Bitbucket) and run your project’s test suite every time a new commit is made.

These services also offer facilities to run your project’s test suite on multiple versions of Python, giving rapid feedback about whether the code will work, without the developer having to perform such tests themselves.

Wikipedia has an extensive comparison of many continuous-integration systems. There are two hosted services which when used in conjunction provide automated testing across Linux, Mac and Windows:

  • Travis CI provides both a Linux and a Mac OSX environment. The Linux environment is Ubuntu 12.04 LTS Server Edition 64 bit while the OSX is 10.9.2 at the time of writing.
  • Appveyor provides a Windows environment (Windows Server 2012).
TODO Either link to or provide example .yml files for these two

TODO How do we keep the Travis Linux and OSX versions up-to-date in this

Both Travis CI and Appveyor require a YAML-formatted file as specification for the instructions for testing. If any tests fail, the output log for that specific configuration can be inspected.

For Python projects that are intended to be deployed on both Python 2 and 3 with a single-source strategy, there are a number of options.

Tools for single-source Python packages

six is a tool developed by Benjamin Peterson for wrapping over the differences between Python 2 and Python 3. The six package has enjoyed widespread use and may be regarded as a reliable way to write a single-source python module that can be use in both Python 2 and 3. The six module can be used from as early as Python 2.5. A tool called modernize, developed by Armin Ronacher, can be used to automatically apply the code modifications provided by six.

Similar to six, python-future is a package that provides a compatibility layer between Python 2 and Python 3 source code; however, unlike six, this package aims to provide interoperability between Python 2 and Python 3 with a language syntax that matches one of the two Python versions: one may use

  • a Python 2 (by syntax) module in a Python 3 project.
  • a Python 3 (by syntax) module in a Python 2 project.

Because of the bi-directionality, python-future offers a pathway to converting a Python 2 package to Python 3 syntax module-by-module. However, in contrast to six, python-future is supported only from Python 2.6. Similar to modernize for six, python-future comes with two scripts called futurize and pasteurize that can be applied to either a Python 2 module or a Python 3 module respectively.

Use of six or python-future adds an additional runtime dependency to your package: with python-future, the futurize script can be called with the --stage1 option to apply only the changes that Python 2.6+ already provides for forward-compatibility to Python 3. Any remaining compatibility problems would require manual changes.

What’s in Which Python?

Ned Batchelder provides a list of changes in each Python release for Python 2 and separately for Python 3. These lists may be used to check whether any changes between Python versions may affect your package.

TODO These lists should be reproduced here (with permission).

TODO The py3 list should be updated to include 3.4