Porting to Python 3 - The Book Site

Porting with 2to3

Although it’s perfectly possible to just run your Python 2 code under Python 3 and fix each problem as it turns up, this quickly becomes very tedious. You need to change every print statement to a print() function, and you need to change every except Exception, e to use the new except Exception as e syntax. These changes are tricky to do in a search and replace and stops being exciting very quickly. Luckily we can use 2to3 to do most of these boring changes automatically.

Using 2to3

2to3 is made up of several parts. The core is a lib2to3, a package that contains support for refactoring Python code. It can analyze the code and build up a parse tree describing the structure of the code. You can then modify this parse tree and from it generate new code. Also in lib2to3 is a framework for “fixers”, which are modules that make specific refactorings, such as changing the print statement into a print() function. The set of standard fixers that enables 2to3 to convert Python 2 code to Python 3 code are located in lib2to3.fixes. Lastly is the 2to3 script itself, which you run to do the conversion.

Running 2to3 is very simple. You give a file or directory as parameter and it will convert the file or look through the directory for Python files and convert them. 2to3 will print out all the changes to the output, but unless you include the -w flag it will not write the changes to the files. It writes the new file to the same file name as the original one and saves the old one with a .bak extension, so if you want to keep the original file name without changes you must first copy the files and then convert the copy.

If you have doctests in your Python files you also need to run 2to3 a second time with -d to convert the doctests and if you have text files that are doctests you need to run 2to3 on those explicitly. A complete conversion can therefore look something like this:

$ 2to3 -w .
$ 2to3 -w -d .
$ 2to3 -w -d src/mypackage/README.txt
$ 2to3 -w -d src/mypackage/tests/*.txt

Under Linux and OS X the 2to3 script is installed in the same folder as the Python executable. Under Windows it in installed as 2to3.py in the Tools\Scripts folder in your Python installation, and you have to give the full path:

C:\projects\mypackage> C:\Python3.3\Tools\Scripts\2to3.py -w .

If you run 2to3 often you might want to add the Scripts directory to your system path.

Explicit fixers

By default, the conversion uses all fixers in the lib2to3.fixers package except buffer, idioms, set_literal and ws_comma. These will only be used if specifically added to the command line. In this case you also need to specify the default set of fixers, called all:

$ 2to3 -f all -f buffer .

The buffer fixer will replace all use of the buffer type with a memoryview type. The buffer type is gone in Python 3 and the memoryview type is not completely compatible. So the buffer fixer is not included by default as you might have to make manual changes as well.

The other three fixers will make more stylistic changes and are as such not really necessary.

The idioms fixer will update some outdated idioms. It will change type(x) == SomeType and other type-tests to using isinstance(), it will change the old style while 1: used in Python 1 into while True: and it will change some usage of .sort() into sorted() (See Use sorted() instead of .sort().)

The set_literal fixer will change calls to the set() constructor to use the new set literal. See Set literals.

The ws_comma fixer will fix up the whitespace around commas and colons in the code.

It is possible to write your own fixers, although it is highly unlikely that you would need to. For more information on that, see Extending 2to3 with your own fixers.

You can also exclude some fixers while still running the default all set of fixers:

$ 2to3 -x print .

If you don’t intend to continue to support Python 2, that’s all you need to know about 2to3. You only need to run it once and then comes the fun part, fixing the migration problems, which is discussed in Common migration problems.

Distributing packages

When you write Python modules or packages that are used by other developers you probably want to continue to support Python 2. Then you need to make sure that Python 3 users get the Python 3 version and Python 2 users get the Python 2 version of your package. This can be as simple as documenting on the download page, if you host your packages yourself.

Most packages for general use use distutils and are uploaded to the CheeseShop[1], from where they are often installed with tools like easy_install or pip. These tools will download the latest version of the package and install it, and if you have both Python 2 and Python 3 packages uploaded to CheeseShop, many of the users will then get the wrong version and will be unable to install your package.

One common solution for this is to have two separate package names, like mymodule and mymodule3, but then you have two packages to maintain and two releases to make. A better solution is to include both source trees in one distribution archive, for example under src2 for Python 2 and src3 under Python 3. You can then in your setup.py select which source tree should be installed depending on the Python version:

import sys
from distutils.core import setup

if sys.version < '3':
      package_dir = {'': 'src2'}
else:
      package_dir = {'': 'src3'}

setup(name='foo',
      version='1.0',
      package_dir = package_dir,
      )

This way all users of your module package will download the same distribution and the install script will install the correct version of the code. Your setup.py needs to run under both Python 2 and Python 3 for this to work, which is usually not a problem. See Supporting Python 2 and 3 without 2to3 conversion for more help on how to do that.

If you have a very complex setup.py you might want to have one for each version of Python, one called setup2.py for Python 2 and one called setup3.py for Python 3. You can then make a setup.py that selects the correct setup-file depending on Python version:

import sys
if sys.version < '3':
    import setup2
else:
    import setup3

Running 2to3 on install

The official way to support both Python 2 and Python 3 is to maintain the code in a version for Python 2 and convert it to Python 3 with the 2to3 tool. If you are doing this you can simplify your distribution by running the conversion during install. That way you don’t have to have separate packages or even two copies of the code in your package distribution.

Distutils supports this with a custom build command. Replace the build_py command class with build_py_2to3 under Python 3:

try:
   from distutils.command.build_py import build_py_2to3 \
        as build_py
except ImportError:
   from distutils.command.build_py import build_py

setup(
   ...
   cmdclass = {'build_py': build_py}
   )

However, if you want to use this solution, you probably want to switch from Distutils to Distribute, that extends this concept further and integrates 2to3 tighter into the development process.

Supporting multiple versions of Python with Distribute

If you are using 2to3 to support both Python 2 and Python 3 you will find Distribute[2] very helpful. It is a Distutils extension that is a Python 3 compatible fork of Phillip J. Eby’s popular Setuptools package. Distribute adds the same 2to3 integration to the build command as Distutils does, so it will solve the distribution problems for you, but it also will help you during the development and testing of your package.

When you use 2to3 to support both Python 2 and Python 3 you need to run 2to3 every time you have made a change, before running the tests under Python 3. Distribute integrates this process into its test command, which means that any files you have updated will be copied to a build directory and converted with 2to3 before the tests are run in that build directory, all by just one command. After you have made a change to your code, you just run python setup.py test for each version of Python you need to support to make sure that the tests run. This makes for a comfortable environment to do Python 3 porting while continuing to support Python 2.

To install Distribute you need run the Distribute setup script from http://python-distribute.org/distribute_setup.py. You then run distribute_setup.py with all Python versions where you want Distribute installed. Distribute is mainly compatible with Setuptools, so you can use Setuptools under Python 2 instead of Distribute but it’s probably better to be consistent and use Distribute under Python 2 as well.

If you are using Distutils or Setuptools to install your software you already have a setup.py. To switch from Setuptools to Distribute you don’t have to do anything. To switch from Distutils to Distribute you need to change where you import the setup() function from. In Distutils you import from distutils.core while in Setuptools and Distribute you import from setuptools.

If you don’t have a setup.py you will have to create one. A typical example of a setup.py would look something like this:

from setuptools import setup, find_packages

readme = open('docs/README.txt', 'rt').read()
changes = open('docs/CHANGES.txt', 'rt').read()

setup(name='Porting to Python 3 examples',
      version="1.0",
      description="An example project for Porting to Python 3",
      long_description=readme + '\n' + changes,
      classifiers=[
          "Programming Language :: Python :: 2",
          "Topic :: Software Development :: Documentation"],
      keywords='python3 porting documentation examples',
      author='Lennart Regebro',
      author_email='regebro@gmail.com',
      license='GPL',
      packages=find_packages(exclude=['ez_setup']),
      include_package_data=True)

Explaining all the intricacies and possibilities in Distribute is outside the scope of this book. The full documentation for Distribute is on http://packages.python.org/distribute.

Running tests with Distribute

Once you have Distribute set up to package your module you need to use Distribute to run your tests. You can tell Distribute where to look for tests to run by adding the parameter test_suite to the setup() call. It can either specify a module to run, a test class to run, or a function that returns a TestSuite object to run. Often you can set it to the same as the base package name. That will make Distribute look in the package for tests to run. If you also have separate files that should be run as DocTests then Distribute will not find them automatically. In those cases it’s easiest to make a function that returns a TestSuite with all the tests.

import unittest, doctest, StringIO

class TestCase1(unittest.TestCase):
    
    def test_2to3(self):
        assert True
    
def test_suite():
    suite = unittest.makeSuite(TestCase1)
    return suite

We then specify that function in the setup():

from setuptools import setup, find_packages

readme = open('docs/README.txt', 'rt').read()
changes = open('docs/CHANGES.txt', 'rt').read()

setup(name='Porting to Python 3 examples',
      version="1.0",
      description="An example project for Porting to Python 3",
      long_description=readme + '\n' + changes,
      classifiers=[
          "Programming Language :: Python :: 2",
          "Topic :: Software Development :: Documentation"],
      keywords='python3 porting documentation examples',
      author='Lennart Regebro',
      author_email='regebro@gmail.com',
      license='GPL',
      packages=find_packages(exclude=['ez_setup']),
      include_package_data=True,
      test_suite='py3example.tests.test_suite')

You can now run your tests with python setup.py test.

Running 2to3 with Distribute

Once you have the tests running under Python 2, you can add the use_2to3 keyword options to setup() and start running the tests with Python 3. Also add "Programming Language :: Python :: 3" to the list of classifiers. This tells the CheeseShop and your users that you support Python 3.

from setuptools import setup, find_packages

readme = open('docs/README.txt', 'rt').read()
changes = open('docs/CHANGES.txt', 'rt').read()

setup(name='Porting to Python 3 examples',
      version="1.0",
      description="An example project for Porting to Python 3",
      long_description=readme + '\n' + changes,
      classifiers=[
          "Programming Language :: Python :: 2",
          "Programming Language :: Python :: 3",
          "Topic :: Software Development :: Documentation"],
      keywords='python3 porting documentation examples',
      author='Lennart Regebro',
      author_email='regebro@gmail.com',
      license='GPL',
      packages=find_packages(exclude=['ez_setup']),
      include_package_data=True,
      test_suite='py3example.tests.test_suite',
      use_2to3=True)

Under Python 3, the test command will now first copy the files to a build directory and run 2to3 on them. It will then run the tests from the build directory. Under Python 2, the use_2to3 option will be ignored.

Distribute will convert all Python files and also all doctests in Python files. However, if you have doctests located in separate text files, these will not automatically be converted. By adding them to the convert_2to3_doctests option Distribute will convert them as well.

To use additional fixers, the parameter use_2to3_fixers can be set to a list of names of packages containing fixers. This can be used both for the explicit fixers included in 2to3 and external fixers, such as the fixers needed if you use the Zope Component Architecture.

from setuptools import setup, find_packages

readme = open('docs/README.txt', 'rt').read()
changes = open('docs/CHANGES.txt', 'rt').read()

setup(name='Porting to Python 3 examples',
      version="1.0",
      description="An example project for Porting to Python 3",
      long_description=readme + '\n' + changes,
      classifiers=[
          "Programming Language :: Python :: 2",
          "Programming Language :: Python :: 3",
          "Topic :: Software Development :: Documentation"],
      keywords='python3 porting documentation examples',
      author='Lennart Regebro',
      author_email='regebro@gmail.com',
      license='GPL',
      packages=find_packages(exclude=['ez_setup']),
      include_package_data=True,
      test_suite='py3example.tests.test_suite',
      use_2to3=True,
      convert_2to3_doctests=['doc/README.txt'],
      install_requires=['zope.fixers'],
      use_2to3_fixers=['zope.fixers'])

Attention

When you make changes to setup.py, this may change which files get converted. The conversion process will not know which files was converted during the last run, so it doesn’t know that a file which during the last run was just copied now should be copied and converted. Therefore you often have to delete the whole build directory after making changes to setup.py.

You should now be ready to try to run the tests under Python 3 by running python3 setup.py test. Most likely some of the tests will fail, but as long as the 2to3 process works and the tests run, even if they fail, then you have come a long way towards porting to Python 3. It’s now time to look into fixing those failing tests. Which leads us into discussing the common migration problems.

Footnotes

[1]http://pypi.python.org/
[2]http://pypi.python.org/pypi/distribute