We do not discourage authors to release software on Python 2. While this guide is mostly written with the assumption that software are going to stop Python 2 support, it does perfectly apply to a package that wishes to not support Python 3, or is stopping support for any minor version.

This page gathers information and links to resources allowing a library to stop supporting an older version of Python without causing too much disruption for users who haven’t upgraded to this new version.

Whether you are a user or a developer, being aware of the issue listed here – at least the main points – should ease lots of the pain.

Too long, did not read:

  • Help and encourage users to install pip 9.0+
  • Help and encourage users to install setuptools 24.3+
  • As maintainer, use the new setup(..., python_requires='>=3.4') option.
  • Use pip install [-e] . and do not invoke setup.py directly.
  • Fail early at install time if user is on Python 2.

We gave a talk at PyCon 2017 (slides), and at PyBay (slides).

The problem

Up until December 2016 it was hard to publish a new major version of a library that changed requirements in Python version and mark it as such so that a user’s system will not try to upgrade said library.

With the recent changes in Python packaging this is now possible.

As an example let’s look at a non-existent fictitious library.

  • fictitious 1.1, 1.2, 1.3, 1.4 are compatible Python 2.7 and 3.3+
  • fictitious 2.0 has been released and is python 3.4+ only.

As a Python 2.7 user, if I don’t pay attention, or if the library is not correctly tagged, there can be issues when you try to update the library:

$ python -c 'import fictitious; print(fictitious.__version__)'
1.3.2
$ pip install fictitious --upgrade

Either my system will install 2.0, which will not work – the worst case scenario – or fail to install, in which case I will not get the critical 1.4 upgrade.

As a user

Install pip 9.0

If you are already a Python 3 user, you should not encounter a lot of disruption. Please still check that the libraries you use follow best practices not to break for Python 2 users. Python is a community regardless of which python version you have to (or decided to) run, making sure that everything works makes the community strong.

Make sure you have Pip ≥ 9.0, this is especially important if you have Python 2 installations. Having pip 9.0+ is not a guarantee of a flawless upgrade. But pip 9.0+ does have a number of safety check not available in previous versions.

Having a version of pip < 9.0 can lead your system to try to upgrade to non-compatible versions of Python packages even if these are marked as non-compatible.

Help as many other users as possible to install pip ≥ 9.0. For the transition, it is the slowest part of the ecosystem to update, and is the only piece that requires action of all Python users.

The simplest way to make sure all is up to date is to run the following for each installation of Python:

$ pip install --upgrade setuptools pip

This will install the latest version of pip and setuptools.

You can issue the following to see the version of pip:

$ pip --version
9.0.0

All good.

Setuptools

If you are on a system for which no wheel is available, pip will try to install a source distribution (aka sdist).

Installing an sdist will require setuptools, so make sure you have setuptools ≥ 24.2.0 or building Python 3-only libraries will likely fail. In particular if library authors have taken time to mark their library as Python 3 only, the python_requires argument to setup() may not be recognized and installation will fail.

Use the following to check your setuptools version :

$ python -c 'import setuptools; print(setuptools.__version__)'
24.2.0

Again make sure to upgrade pip and setuptools to make sure you have an up to date system:

$ pip install --upgrade setuptools pip

Local package index

If you are using a custom local package index, for example if you are working at a company with private packages, make sure it correctly implements pep-503 to let pip know about the python_requires field. This mostly mean that the HTML you are exposing should get a data-python-requires data attribute with the (html escaped) version specifier.

The state of PyPI

Warehouse and Legacy PyPI have received various patches to ensure they support this new functionality.

Preparing your library

As a library author one of the most important factors in a smooth transition is planning and communication, letting your user base know in advance that the transition is happening and what step to take is critical for a transition.

For your library code here the steps you need to take to ensure that installation will fail in the least number of cases:

You need to release your package’s new version with setuptools version 24.2.0 or above. You can also use one of the alternate package managers that can set the Requires-Python metadata field. Without this, pip 9.0 will try to install a non-compatible version of your software on Python 2. This version of setuptools is recent (July 20, 2016) and this is all possible thanks to the work of Xavier Fernandez

Add the following to your setup.py

setup(
   ...
   python_requires='>=3.3'
   ...
)

Change >=3.3 accordingly depending on what version your library decides to support. In particular you can use >=2.6 or >=3.5 ! Note that this also support the compatible with syntax: ~=2.5 (meaning, >=2.5 and <3).

This will make PyPI aware that your package is Python 3.3+ only, and allow pip to be made aware of this.

Thus as long as your user have recent enough versions of pip and setuptools they will get the right version of your library.

Unit testing and documentation

It is recommended not to invoke setup.py directly either with install or develop subcommands. These may not correctly resolve dependencies, and can install incompatible versions of dependencies. Please recommend and use pip install . and pip install -e . for regular and developer installs, respectively.

Check in scripts and documentation that the correct installation command is used.

Recommended mitigations

These are not mandatory but should make the transition seamless by warning your users early enough and providing useful error messages.

Runtime warning on master

Add a warning at runtime that triggers early on master (before switching to Python 3 only)

import warnings
import sys
if sys.version_info < (3,):
    warnings.warn('You are using master of `Frobulator` with Python 2. '
                  'Frobulator will soon be Python 3 only. '
                  'See this issue to know more.',
                  UserWarning)

Your Python 2 users will have a chance to upgrade, or get off master, (for example on the LTS branch).

Fail early at import time

Add an error early through import at runtime with a clear error message, leave the early import compatible Python 2 as users will not feel welcomed with a useless SyntaxError due to their Python 2 usage. Don’t hesitate to use multi-line strings in error messages.

Error at import time will happen on systems with old version of pip and setuptools. Keep in mind that saying the package is Python 3 only is not a lot more helpful than a SyntaxError. The most reasonable reason would be out-of-date pip and setuptools:

import sys

if sys.version_info < (3,):
    raise ImportError(
    """You are running Frobulator 6.0 on Python 2

Frobulator 6.0 and above are no longer compatible with Python 2, and you still 
ended up with this version installed. That's unfortunate; sorry about that. 
It should not have happened. Make sure you have pip >= 9.0 to avoid this kind 
of issue, as well as setuptools >= 24.2:

 $ pip install pip setuptools --upgrade

Your choices:

- Upgrade to Python 3.

- Install an older version of Frobulator:

 $ pip install 'frobulator<6.0'

It would be great if you can figure out how this version ended up being
installed, and try to check how to prevent that for future users.

See the following URL for more up-to-date information:

https://i.am.an/url

""")

Watch out for beta releases

Make sure your version number matches PEP 440 or you will get surprises during beta, in particular as the sdist and wheel will appear as being different versions (the sdist (during beta/rc/post) can appear with a greater version number than wheels). Pip thus will try to install the sdist instead of the wheel, which has more chance of failing, in particular with pre-24.2 versions of setuptools.

The regular expression to check for validity of pep440 can be found below:

^
([1-9]\\d*!)?
(0|[1-9]\\d*)
(\\.(0|[1-9]\\d*))*
((a|b|rc)(0|[1-9]\\d*))?
(\\.post(0|[1-9]\\d*))?
(\\.dev(0|[1-9]\\d*))?

Fail early in setup.py

Leave setup.py python 2 compatible and fail early. If you detect Python 2 raise a clear error message and ask the user to make sure they have pip > 9.0 (or migrate to Python 3). You can (try to) conditionally import pip and check for its version but this might not be the same pip. Failing early is important to make sure the Python installation does not install an incompatible version. Otherwise user code can fail at runtime arbitrarily later in the future, which can be a difficult to debug and fix. Get inspiration from the message of failure at runtime, and adapt for installation time.

Here is for a simple version of how IPython handle old versions of Python and Pip that still try to install IPython 7.0+ on Python < (3,4).

if sys.version_info < (3, 4):

    error = """
    IPython 7.0+ supports Python 3.4 and above.
    When using Python 2.7, please install IPython 5.x LTS Long Term Support version.
    Python 3.3 was supported up to IPython 6.x.

    See IPython `README.rst` file for more information:

    https://github.com/ipython/ipython/blob/master/README.rst

    Python {py} detected.

    Try upgrading pip and retry.
    """.format(py='.'.join([str(v) for v in sys.version_info[:3]]))

    print(error, file=sys.stderr)
    sys.exit(1)

You can look at the full check that attempt to detect which version of pip is in used.

Upload with Twine.

You must upload your package with twine and NOT with setup.py upload. If you have an old version of setuptools or distutils, even if python_requires is set, metadata may not be uploaded correctly. See This issue on PyPI repository

Fix dependant libraries

If you control dependant packages, Make sure to include conditional dependencies depending on the version of Python.

Non-recommended mitigations

This is a collection of “mitigation” or “solutions” you will find on the web and that you will hear about. This is an attempt to acknowledge them, and explain why they can’t work and what are their drawbacks before you attempt to implement them.

Use a meta-package

It is possible to release a meta-package that has virtually no code and relies on a conditional dependency to install its actual core code on the user system. For example, Frob-6.0 could be a meta-package which depends on Frob-real-py2 on Python < 3.0, and Frob-real-py3 on Python ≥ 3.4. While this approach is doable this can make imports confusing.

Depend on setuptools

You can mark your library as dependent on setuptools greater than 24.3 as this will ensure that during the next upgrade (when the packages drop python 2 support) will have the right version of setuptools.

Of course regardless of all the care you will take for your library to no break and to install only on python 2, you will likely have cases where it will still end up being installed on incompatible versions of Python. Simply because users upgrade rarely and only an old version of pip or setuptools is enough to make the update process broken.

Plus setuptools is rarely an actual dependency of your project but a requirement to build wheels.

Multiple sdist files

Pip (used to) support a “feature” where a sdist ending in -pyX.Y.tar.gz would only be seen as compatible on Python X.Y, thus it used to be possible to publish multiple sdist of a package targeting various python version.

It is not possible anymore to upload multiple sdist files on PyPI, so this solution is no longer tenable.

Wheel only?

Releasing a package only using wheels for a given python version is doable, but this will break downstream packages that may require the original source to reproduce their build.

Why all this?!?

You might wonder why all this, it’s 2018 already, so how come this is still an issue? Python 3 has been out for 9+ years now!

Well there are many reasons for this. First of all, this issue mostly affects libraries that are currently python 2 and Python 3 compatible at the same time. Many libraries have transitioned from Python 2-only to Python 2 + 3. And the issue of transitioning to Python 3 only is relatively recent. Technically it can also apply to libraries that are only stopping support for 2.6, or are even already Python 3 only, but are starting to stop supporting earlier versions of Python (for example a library releasing a Python 3.4+ only version).

Python 3.3 was released at the end of 2012, and was the first version to support (again) u as a prefix for Unicode string. It was one of the first minor versions of Python 3 that saw a majority of single-source projects working both on Python 2 and Python 3. These are the projects that will likely be affected by this issue.

The introduction of Python 3 was chaotic; there are still strong arguments in both the Python 2 and Python 3 camps. Regardless of what side you take, the ones suffering the most from this are users (starting with the fact that inevitably some libraries will stop supporting for Python 2 and release Python 3 only library). Inevitably, some systems and people will not be upgraded to Python 3, so this document hopefully helps to ensure that users get the least breakage as possible and what are the best practices are to follow.