Porting to Python 3

Django 1.5 is the first version of Django to support Python 3. The same code runs both on Python 2 (≥ 2.6.5) and Python 3 (≥ 3.2), thanks to the six compatibility layer.

This document is primarily targeted at authors of pluggable applications who want to support both Python 2 and 3. It also describes guidelines that apply to Django’s code.

Philosophy

This document assumes that you are familiar with the changes between Python 2 and Python 3. If you aren’t, read Python’s official porting guide first. Refreshing your knowledge of unicode handling on Python 2 and 3 will help; the Pragmatic Unicode presentation is a good resource.

Django uses the Python 2/3 Compatible Source strategy. Of course, you’re free to chose another strategy for your own code, especially if you don’t need to stay compatible with Python 2. But authors of pluggable applications are encouraged to use the same porting strategy as Django itself.

Writing compatible code is much easier if you target Python ≥ 2.6. Django 1.5 introduces compatibility tools such as django.utils.six, which is a customized version of the six module. For convenience, forwards-compatible aliases were introduced in Django 1.4.2. If your application takes advantage of these tools, it will require Django ≥ 1.4.2.

Obviously, writing compatible source code adds some overhead, and that can cause frustration. Django’s developers have found that attempting to write Python 3 code that’s compatible with Python 2 is much more rewarding than the opposite. Not only does that make your code more future-proof, but Python 3’s advantages (like the saner string handling) start shining quickly. Dealing with Python 2 becomes a backwards compatibility requirement, and we as developers are used to dealing with such constraints.

Porting tools provided by Django are inspired by this philosophy, and it’s reflected throughout this guide.

Porting tips

Unicode literals

This step consists in:

  • Adding from __future__ import unicode_literals at the top of your Python modules – it’s best to put it in each and every module, otherwise you’ll keep checking the top of your files to see which mode is in effect;
  • Removing the u prefix before unicode strings;
  • Adding a b prefix before bytestrings.

Performing these changes systematically guarantees backwards compatibility.

However, Django applications generally don’t need bytestrings, since Django only exposes unicode interfaces to the programmer. Python 3 discourages using bytestrings, except for binary data or byte-oriented interfaces. Python 2 makes bytestrings and unicode strings effectively interchangeable, as long as they only contain ASCII data. Take advantage of this to use unicode strings wherever possible and avoid the b prefixes.

Note

Python 2’s u prefix is a syntax error in Python 3.2 but it will be allowed again in Python 3.3 thanks to PEP 414. Thus, this transformation is optional if you target Python ≥ 3.3. It’s still recommended, per the “write Python 3 code” philosophy.

String handling

Python 2’s unicode type was renamed str in Python 3, str() was renamed bytes, and basestring disappeared. six provides tools to deal with these changes.

Django also contains several string related classes and functions in the django.utils.encoding and django.utils.safestring modules. Their names used the words str, which doesn’t mean the same thing in Python 2 and Python 3, and unicode, which doesn’t exist in Python 3. In order to avoid ambiguity and confusion these concepts were renamed bytes and text.

Here are the name changes in django.utils.encoding:

Old name New name
smart_str smart_bytes
smart_unicode smart_text
force_unicode force_text

For backwards compatibility, the old names still work on Python 2. Under Python 3, smart_str is an alias for smart_text.

For forwards compatibility, the new names work as of Django 1.4.2.

Note

django.utils.encoding was deeply refactored in Django 1.5 to provide a more consistent API. Check its documentation for more information.

django.utils.safestring is mostly used via the mark_safe() and mark_for_escaping() functions, which didn’t change. In case you’re using the internals, here are the name changes:

Old name New name
EscapeString EscapeBytes
EscapeUnicode EscapeText
SafeString SafeBytes
SafeUnicode SafeText

For backwards compatibility, the old names still work on Python 2. Under Python 3, EscapeString and SafeString are aliases for EscapeText and SafeText respectively.

For forwards compatibility, the new names work as of Django 1.4.2.

__str__() and __unicode__() methods

In Python 2, the object model specifies __str__() and  __unicode__() methods. If these methods exist, they must return str (bytes) and unicode (text) respectively.

The print statement and the str built-in call __str__() to determine the human-readable representation of an object. The unicode built-in calls  __unicode__() if it exists, and otherwise falls back to __str__() and decodes the result with the system encoding. Conversely, the Model base class automatically derives __str__() from  __unicode__() by encoding to UTF-8.

In Python 3, there’s simply __str__(), which must return str (text).

(It is also possible to define __bytes__(), but Django applications have little use for that method, because they hardly ever deal with bytes.)

Django provides a simple way to define __str__() and  __unicode__() methods that work on Python 2 and 3: you must define a __str__() method returning text and to apply the python_2_unicode_compatible() decorator.

On Python 3, the decorator is a no-op. On Python 2, it defines appropriate  __unicode__() and __str__() methods (replacing the original __str__() method in the process). Here’s an example:

from __future__ import unicode_literals
from django.utils.encoding import python_2_unicode_compatible

@python_2_unicode_compatible
class MyClass(object):
    def __str__(self):
        return "Instance of my class"

This technique is the best match for Django’s porting philosophy.

For forwards compatibility, this decorator is available as of Django 1.4.2.

Finally, note that __repr__() must return a str on all versions of Python.

dict and dict-like classes

dict.keys(), dict.items() and dict.values() return lists in Python 2 and iterators in Python 3. QueryDict and the dict-like classes defined in django.utils.datastructures behave likewise in Python 3.

six provides compatibility functions to work around this change: iterkeys(), iteritems(), and itervalues(). It also contains an undocumented iterlists function that works well for django.utils.datastructures.MultiValueDict and its subclasses.

HttpRequest and HttpResponse objects

According to PEP 3333:

  • headers are always str objects,
  • input and output streams are always bytes objects.

Specifically, HttpResponse.content contains bytes, which may become an issue if you compare it with a str in your tests. The preferred solution is to rely on assertContains() and assertNotContains(). These methods accept a response and a unicode string as arguments.

Coding guidelines

The following guidelines are enforced in Django’s source code. They’re also recommended for third-party applications that follow the same porting strategy.

Syntax requirements

Unicode

In Python 3, all strings are considered Unicode by default. The unicode type from Python 2 is called str in Python 3, and str becomes bytes.

You mustn’t use the u prefix before a unicode string literal because it’s a syntax error in Python 3.2. You must prefix byte strings with b.

In order to enable the same behavior in Python 2, every module must import unicode_literals from __future__:

from __future__ import unicode_literals

my_string = "This is an unicode literal"
my_bytestring = b"This is a bytestring"

If you need a byte string literal under Python 2 and a unicode string literal under Python 3, use the str builtin:

str('my string')

In Python 3, there aren’t any automatic conversions between str and bytes, and the codecs module became more strict. str.encode() always returns bytes, and bytes.decode always returns str. As a consequence, the following pattern is sometimes necessary:

value = value.encode('ascii', 'ignore').decode('ascii')

Be cautious if you have to index bytestrings.

Exceptions

When you capture exceptions, use the as keyword:

try:
    ...
except MyException as exc:
    ...

This older syntax was removed in Python 3:

try:
    ...
except MyException, exc:    # Don't do that!
    ...

The syntax to reraise an exception with a different traceback also changed. Use six.reraise().

Magic methods

Use the patterns below to handle magic methods renamed in Python 3.

Iterators

class MyIterator(six.Iterator):
    def __iter__(self):
        return self             # implement some logic here

    def __next__(self):
        raise StopIteration     # implement some logic here

Boolean evaluation

class MyBoolean(object):

    def __bool__(self):
        return True             # implement some logic here

    def __nonzero__(self):      # Python 2 compatibility
        return type(self).__bool__(self)

Division

class MyDivisible(object):

    def __truediv__(self, other):
        return self / other     # implement some logic here

    def __div__(self, other):   # Python 2 compatibility
        return type(self).__truediv__(self, other)

    def __itruediv__(self, other):
        return self // other    # implement some logic here

    def __idiv__(self, other):  # Python 2 compatibility
        return type(self).__itruediv__(self, other)

Special methods are looked up on the class and not on the instance to reflect the behavior of the Python interpreter.

Writing compatible code with six

six is the canonical compatibility library for supporting Python 2 and 3 in a single codebase. Read its documentation!

A customized version of six is bundled with Django as of version 1.4.2. You can import it as django.utils.six.

Here are the most common changes required to write compatible code.

String handling

The basestring and unicode types were removed in Python 3, and the meaning of str changed. To test these types, use the following idioms:

isinstance(myvalue, six.string_types)       # replacement for basestring
isinstance(myvalue, six.text_type)          # replacement for unicode
isinstance(myvalue, bytes)                  # replacement for str

Python ≥ 2.6 provides bytes as an alias for str, so you don’t need six.binary_type.

long

The long type no longer exists in Python 3. 1L is a syntax error. Use six.integer_types check if a value is an integer or a long:

isinstance(myvalue, six.integer_types)      # replacement for (int, long)

xrange

If you use xrange on Python 2, import six.moves.range and use that instead. You can also import six.moves.xrange (it’s equivalent to six.moves.range) but the first technique allows you to simply drop the import when dropping support for Python 2.

Moved modules

Some modules were renamed in Python 3. The django.utils.six.moves module (based on the six.moves module) provides a compatible location to import them.

PY2

If you need different code in Python 2 and Python 3, check six.PY2:

if six.PY2:
    # compatibility code for Python 2

This is a last resort solution when six doesn’t provide an appropriate function.

Django customized version of six

The version of six bundled with Django (django.utils.six) includes a few customizations for internal use only.