Armin Ronacher's Thoughts and Writings

Porting to Python 3 — A Guide

written on Thursday, February 11, 2010

The latest Jinja 2 release came with basic support for Python 3. It was surprisingly painless to port the application over but it did require a substantial amount of tweaks and code changes in order to get it running. For everyone else out there who is interested in getting started, I decided to share my experiences:

Changing APIs

Before you start porting the library you have to decide how interfaces will behave in Python 3. The biggest issue here is obviously unicode, but there are others as well. I would say there are four kinds of libraries you might encounter regarding string behavior in Python 2: There are the libraries that only accept unicode and only output unicode, there are those that only accept byte-strings and output byte-strings but operate on textual data, there are the libraries that operate on either or and what has been fed into it, comes out of it and there are libraries that operate either on unicode or byte-strings and also accept the other type as long as it's a subset of the default encoding (ASCII).

First you have to find out what your library does, what it is supposed to do, and how you want to deal with that in Python 3. Because byte-strings no longer exist in Python 3 and were replaced by a bytes object that works similar, but has an incompatible API it is very unlikely that your code will be able to support both in the future (or that it is something you would desire).

Byte-Based Libraries

This is might the most tricky one if you are aiming for Python 2.5 support or lower and you are operating on bytes directly. The issue is that the way you operate on bytes changed fundamentally from Python 2.x to 3.x and 2to3 is not really able to pick it up. Worse, it will try convert all your bytestring literals to unicode! The official support is as far as I know, to explicitly prefix the byte strings in the 2.x code with a leading b to indicate bytes. Unfortunately that means no support for 2.x. I am not completely sure what to do in that situation, but at least I found a way to trick python to operate on bytes: if you have code like this:

magic = 'M23\x01'

And you want to ensure it does not end up being a str in 3.x, add a dummy encode:

magic = u'M23\x01'.encode('iso-8859-1')

The only downside is that the encode happens at runtime, so it will slow down execution a bit.

Text Based Libraries

The second kind of library is a library that operates on text. In 2.x there were multiple ways to implement such libraries and it basically came down to what data type was used internally and what was accepted for input and output. There are the libraries that operate exclusively either on bytestrings or unicode. These are the ones that are the easiest to port, because 2to3 was written with nearly that in mind. If your library was only accepting bytestrings in 2.x it will (after a 2to3 run) only be accepting a Python 3 str type which is unicode based. This works well as long as you do not intend to use some kind of IO in your library. Once you start doing that, you will need to make sure you can somehow specify the encoding to be used when opening files. In that case, make sure you open the file in byte mode (not in text mode!) and do the decoding/encoding yourself. This is the only way your IO code will work the same in both 2.x and 3.x. But more on IO later.

What 2to3 does out of the box is converting calls from unicode to str automatically. Unfortunately it does not change the special __unicode__ method to __str__. You can easily do that in a custom fixer though, so it should be easy to accomplish. If your library however supports both __str__ and __unicode__ you are in a more tricky situation here. Let me show you an example of the kind of classes I deal with in Jinja 2 for example:

class MyObject(object):

    def __init__(self):
        self.value = u'some value'

    def __str__(self):
        return unicode(self).encode('utf-8')

    def __unicode__(self):
        return self.value

The big problem here is that 2to3 will convert it to this:

class MyObject(object):

    def __init__(self):
        self.value = 'some value'

    def __str__(self):
        return str(self).encode('utf-8')

    def __unicode__(self):
        return self.value

If you call str() on your instance now, it will die with a runtime error because it recurses infinitely. Even if it would not recurse, it would try to return a bytes object from the __str__ method because of the encode call. My plan was to write a custom fixer that, if it detects a __str__ that just calls into __unicode__ and encodes, will drop the __str__ method and rename __unicode__ to __str__. Unfortunately the tree you are dealing with in 2to3 does not appear to be designed to removing code so what I do instead of removing the __str__ is just renaming the __unicode__ to __str__ and let Python override the dummy __str__ with the correct one. The fixer I use for that, looks like this:

from lib2to3 import fixer_base
from lib2to3.fixer_util import Name

class FixRenameUnicode(fixer_base.BaseFix):
    PATTERN = r"funcdef< 'def' name='__unicode__' parameters< '(' NAME ')' > any+ >"

    def transform(self, node, results):
        name = results['name']
        name.replace(Name('__str__', prefix=name.prefix))

After conversion with this fixer in place, the class from above will then look like this:

class MyObject(object):

    def __init__(self):
        self.value = 'some value'

    def __str__(self):
        return str(self).encode('utf-8')

    def __str__(self):
        return self.value

But where to put those fixers? Edit 2to3 directly? And do I have to provide two source packages for 2.x and 3.x? This is where distribute comes in.

2to3 through distribute

Distutils itself already has the possibility to run 2to3 for you, but what it cannot do is adding custom fixers without a lot of custom code. distribute on the other hand not gives you built in 2to3 support as a single keyword argument to setup() but can also pass custom fixers to 2to3 which is very helpful. Because these new keyword arguments however would warn if the setup script was executed with setuptools instead of distribute, you should only pass them to the setup function if invoked from Python 3. The setup script then looks like this:

import sys

from setuptools import setup

# if we are running on python 3, enable 2to3 and
# let it use the custom fixers from the custom_fixers
# package.
extra = {}
if sys.version_info >= (3, 0):

    name='Your Library',
        # make sure to use :: Python *and* :: Python :: 3 so
        # that pypi can list the package on the python 3 page
        'Programming Language :: Python',
        'Programming Language :: Python :: 3'
    # make sure to add custom_fixers to the

Now all you have to do is to put the custom 2to3 fixers (written in Python 3!) into the custom_fixers package next to your real library and they will be added automatically. For examples of fixers, look into the lib2to3/fixes package or your Python 3 installation. If you run python3 build it will run 2to3 on your files and put the output into the build folder for you to test.


So in Python 3 there is a completely new input/output system. It is very Java-ish and is able to deal with unicode. The downside is that you either don't have it in 2.x or the implementation is too slow, so what you want to do is to create yourself an abstraction layer.

If your library was unicode based in older Python versions you probably just did or something similar. This still works on 3.x and I strongly recommend doing that, but be sure to open the file in binary mode, otherwise on Python 3 the decode will attempt to decode an already decoded unicode string, which does not make any sense. If you need normalized newlines (windows newlines converted to 'n') you would have to post-process the string by hand, but must applications and libraries are able to deal with any kind of newline anyways.

You could also just create a IO helper module that calls the builtin open on 3.x and on 2.x. Unfortunately has a worse performance than the built in open on 2.x, so you might want to check how you are dealing with files, if a high performance is necessary and so forth. Most of the time, opening the file in binary mode is what you want to do.

If you library was byte based in 2.x and you opened files in the library, instead of just working on open file objects, you will have to change your API slightly in order to take the charset and error mode into account. If you previously had a function like this:

def read_file_contents(filename):
    with open(filename) as f:

You will have to change it to something like this now:

def read_file_contents(filename, charset='utf-8', errors='strict'):
    with open(filename, 'rb') as f:
        return, errors)

And then ensure that you give the user to provide these arguments to the function. This means that whatever calls this, would also have to accept this arguments and so forth. Not everyone is using utf-8, there might be legacy files in iso-8859-1 a user might still want to be able to open. With a proper error handling system, it might even be possible to fall back to another encoding if it does not decode as utf-8 properly.

Last but not least, 3.x StringIO is a "string IO", not something that accepts binary data. If you have a lot of unittests that are dealing with binary data in such objects, you will have to use the io.BytesIO instead. If it does not exist, you are running 2.x, and you can safely fall back to cStringIO.StringIO.


Now the biggest problem I had with switching to 3.x: The unittests. First of all: do not use doctest. There is a doctest converter in 2to3, but it does not give you much. Error messages changed, reprs changed which it cannot properly pick up, nested tracebacks cause a lot of grief and they are hard to debug. I was playing with the idea to write a tool that automatically converts doctests to unittests, but I was too lazy and converted the few I had in my code, to unittests by hand. Furthermore, the few doctests left (used as code examples in the documentation) are only tested if the testsuite is invoked from Python 2.x

Nosetest has 3.x support in a separate branch, py.test comes with 3.x for a while now and the builtin unittest does the trick as well. I personally converted all my Jinja 2 tests to unittest lately. If you are using unittest you can point distribute to your test suite function and it will run the test for you if you write python test. This even runs 2to3 for you if you execute it with Python 3. So very helpful.

Hope that helps you porting your libraries to Python 3. Would love to hear about your experiences, because even if Python 3 did not work out as some of us hoped, it is very important that we continue to port libraries over to 3.x.

This entry was tagged python