Differences between revisions 33 and 34
Revision 33 as of 2013-03-18 16:46:55
Size: 16363
Editor: barry
Revision 34 as of 2013-03-18 16:52:44
Size: 16388
Editor: barry
Deletions are marked like this. Additions are marked like this.
Line 2: Line 2:


Python 3 on Ubuntu

It is a release goal for Ubuntu 14.04 LTS to have only Python 3 on the desktop CD images. We've had blueprints for several Ubuntu releases, the latest of which is a Raring blueprint. There is a more detailed spec for this effort, an official transition tracker and a publicly shared Google docs spreadsheet to track this effort. This is an ambitious effort that will only succeed with help from the greater Ubuntu, Debian, and Python communities. In other words, we need you!

At the bottom of this page, you will find various resources for diving more into aspects of supporting Python 3, from the pure-Python, C extension module, Debian packaging, and other perspectives. The intent of this page is to provide specific guidelines in a quick reference format, so that you only need to go here once you're familiar with the basic concepts and approaches, but need a refresher on specific coding techniques. This is a wiki page, and you are encouraged to contribute, but try to keep your recommendations tightly focused on accomplishing the release goal of Python 3 only on the 14.04 images.

At the bottom of this page, you'll find additional resources that will help you in your porting efforts. Please add to them!

Before you start

Here are recommendations for you to follow before you start porting.

  • Target Python 3.3 and 2.7. If you have to, optionally include 3.2 and 2.6. Ignore anything older than that (it gets much harder to be bilingual the farther back you go).
  • Use a single code base for both Python 2 and 3.
  • Try to avoid the production use of 2to3. However you may find it helpful to use 2to3 in report-only mode to get a good sense of what changes will be necessary for your package.

  • Use the third part six module sparingly (i.e. only if necessary). One good use case is the reraise() method.

  • Even if you can't complete the port to Python 3 because of your dependencies, start by modernize your Python 2, getting it working cleanly with python2.7 -3 first.

  • Clarify your data model: what are bytes (data) and what are strings (text)? If you can't answer this, you are in for a world of hurt (but you probably are already in Python 2).

I cannot overemphasize the last point. Without a clear separation in your mind and data model between bytes and strings, your port will likely be much more painful than it needs to be. This is the biggest distinction between Python 2 and Python 3. Where Python 2 let you be sloppy, with its 8-bit strings that served as both data and ASCII strings, with automatic (but error prone) conversions between 8-bit strings and unicodes, in Python 3 there are only bytes and strings (i.e. unicodes), with no automatic conversion between the two. This is A Good Thing.

Python source

Basic compatibility

Put the following at the top of all your Python files:

from __future__ import absolute_import, print_function, unicode_literals

This turns on three important compatibility flags.

  • Absolute imports are the default in Python 3 [more info]

  • print() is a function in Python 3 [more info]

  • Unadorned string literals are unicode type in Python 3 [more info]

In your code, make these changes:

  • Change all your print statements to use print() functions, and remove all the u'' prefixes from your strings.

  • If you have string literals in your code that represent data, prefix them all with the b'' prefix [more info]

  • Remove all L suffixes from your long integers. [more info]

Commentary: Some folks don't like to import unicode_literals because it has the potential to change your API (e.g. possibly returning unicodes where before it returned 8-bit strings in Python 2). One developer recommends making this change in your tests only, or only doing this if you have exceptionally good test coverage. YMMV.


  • Change usage of xrange() to range(). [more info]

  • When mocking builtins, remember that in Python 3, you need to use the builtins module instead of the __builtin__ module, e.g.

def mock_it(builtin_name):
   name = ('builtins.%s' if sys.version_info >= (3,) else '__builtin__.%s') % builtin_name
   return mock.patch(name)

[more info]

  • Be explicit in the mode you open() files in; don't assume they will be opened in binary or text mode, pass in the t or b flag as appropriate. [more info] (link also contains other good advice for opening files compatibly).


  • Python 2 codecs which do str-to-str conversions (e.g. rot-13) do not work in Python 3. Use this instead:

from codecs import getencoder
encoder = getencoder('rot-13')
rot13string = encoder(mystring)[0]

[more info]

  • Writing unicode safely against LC_ALL=C to stdout:

sys.stdout = io.TextIOWrapper(sys.stdout.detach(), encoding='UTF-8', line_buffering=True)


  • Change all your uses of the dictionary methods iteritems(), iterkeys(), and itervalues() to use the non-iter variety, e.g. items(), keys(), and values() respectively. These return dictionary views in Python 3, not concrete lists, so if you need a concrete list, wrap these calls in list() or sorted(). [more info]


  • In your doctest's setUp(), add these globals to your test object's globs so they'll have the same __future___ environment that your code has:

from __future__ import absolute_import, print_function, unicode_literals
def setUp(testobj):
    testobj.globs['absolute_import'] = absolute_import
    testobj.globs['print_function'] = print_function
    testobj.globs['unicode_literals'] = unicode_literals

[more info]

  • Bytes have different reprs in Python 2 and Python 3. This convenience function is used to print bytes objects in cross-compatible ways:

def print_bytes(obj)
    if bytes is not str:
        obj = repr(obj)[2:-1]


  • gettext.install() only takes unicode=True in Python 2; this keyword is absent and unnecessary in Python 3 (Python 3 always returns unicodes). A compatible call would be:

kwargs = {}
if sys.version_info >= (3,):
    kwargs['unicode'] = True
gettext.install(domain, LOCALEDIR, **kwargs)


  • Change your iterator classes from providing a next() method to providing a __next__() method. For cross-compatibility, in your class, set next = __next__. [more info]

  • Use itertools.zip_longest() in Python 3, with a conditional import for itertools.izip_longest() in Python 2.


  • Syntax for creating instances with different metaclasses is very different between Python 2 and 3. Use the ability to call type instances as a way to portably create such instances. For example (from the flufl.enum package):

# Define the Enum class using metaclass syntax compatible with both Python 2
# and Python 3.
Enum = EnumMetaclass(str('Enum'), (), {
    '__doc__': 'The public API Enum class.',

Here EnumMetaclass is the metaclass (duh!) and Enum is the class you're creating which has the custom metaclass. You pass in the base classes (of which there are none, hence the empty tuple) and the dictionary of attributes for the class you're creating. The use of str() here is a bit odd, but it's because the enum module uses from __future__ import unicode_literals for Python 2, but in Python 2, class names must be str/bytes. We can't use the b'' prefix though because in Python 3, class names must be unicodes.


  • Python 3 has no operator.isSequenceType(). Use the following code for cross-compatibility.

from collections import Sequence
return isinstance(obj, Sequence)


  • Three argument raise is simply not source compatible between Python 2 and 3 (the Python 2 form is a SyntaxError in Python 3). You'll need to use exec() to work around this, or use six.reraise() (which essentially hides the nasty exec() from you!). [more info]


  • bytes objects in Python 3 have no .format() method. Use concatenation instead.

  • For raw-bytes objects, use the br'' string prefix (rb'' was added to Python 3.3)


  • Consider passing universal_newlines=True to subprocess.Popen() and friends to get text output directly.


  • The implements() method does not work in Python 3. Use the @implementer class decorator instead. [more info]

Python extension modules

  • Define a PY3 macro which you can later #ifdef on for C code which cannot be written portably for both Python 2 and Python 3. [example]

Compatibility macros

  • The PyInt_* functions are gone in Python 3. In your extension module, change all of these to PyLong_* functions, which will work in both versions. [more info]

  • #include <bytesobject.h> and change all PyString_* functions with their PyBytes_* equivalents, changing those that really operate on unicodes to use PyUnicode_* functions. [more info]

  • Instead of explicitly dereferencing ob_type, use the Py_TYPE() macro instead. [more info]

C types

There are lots of differences you need to be aware of when defining types in C extensions. A few important ones:

  • Use PyVarObject_HEAD_INIT() and don't define the tp_size slot [more info]

  • Remove references to Py_TPFLAGS_HAVE_WEAKREFS and Py_TPFLAGS_HAVE_ITER since these are unnecessary (and undefined) in Python 3. If you need to support both Python 2 and 3, you'll need an #ifdef.


  • PyArg_Parse() and friends lack a y code (for bytes objects) in Python 2, so you will have to #ifdef around these.

  • In Python 3, there's no equivalent of the z code for bytes objects (accepting None as well). Write an O& converter.


  • Rewrite these to use PyCapsule instead. If you can drop Python 2.6, there's no need to #ifdef these, since PyCapsule is available in Python 2.7. [example]


  • If you derive new types from builtin C types, e.g. PyBytes, and you want to override the repr in the subclass, you'll find you have a problem with cross-compatibility. In Python 2, the super class's repr will return bytes (a.k.a. 8-bit strings) while in Python 3, they will return unicodes. Python's C API has a little known format code %V which can be used to bridge this gap. Add this macro:

#define REPRV(obj) \
    (PyUnicode_Check(obj) ? (obj) : NULL), \
    (PyUnicode_Check(obj) ? NULL : PyBytes_AS_STRING(obj))

and use it like this:

return PyUnicode_FromFormat("...%V...", REPRV(parent_repr));

[more info]

Third party packages

  • If you're using python-apt, you must port entirely to the 0.8 API. python-apt tolerated this under Python 2, but the old API is compiled out under Python 3. /usr/share/python-apt/migrate-0.8.py may be of some partial help.

  • pyflakes doesn't like it when you define methods twice depending on sys.version. You may want to add some excludes for these.

  • python-libxml2 is currently unavailable for Python 3. Consider using something available in both Python 2 and 3 from the stdlib instead, e.g. xml.etree.cElementTree. If necessary, port to python{,3}-lxml.

  • for gstreamer advice seems to be to use v1.0 which works well with pygi. See the migration guide at Novacut/GStreamer1.0

  • If your application is using OAuth, port your code from the old, unmaintained, OAuth1-only, Python 2-only oauth library (i.e. python-oauth) to the new, maintained, OAuth1 and OAuth2 compatible, Python 2 and Python 3 compatible oauthlib (i.e. python-oauthlib) package. Porting is not so hard, although some of the terminology updates in the OAuth spec since the original oauth was released make it a bit trickier.



  • Why not rely on 2to3? 2to3 is a pretty slow tool so it can impede on the speed with which you develop your code. It can however be used in non-overwrite mode to get a good sense of what needs to be changed in your source code. Generally, a single-source approach is the best for bilingual (i.e. Py2 and Py3) Python apps.

Python/3 (last edited 2016-02-26 01:40:43 by cmawebsite)