What’s New in Python 2.4

Author:A.M. Kuchling

This article explains the new features in Python 2.4.1, released on March 30, 2005.

Python 2.4 is a medium-sized release. It doesn’t introduce as many changes as the radical Python 2.2, but introduces more features than the conservative 2.3 release. The most significant new language features are function decorators and generator expressions; most other changes are to the standard library.

According to the CVS change logs, there were 481 patches applied and 502 bugs fixed between Python 2.3 and 2.4. Both figures are likely to be underestimates.

This article doesn’t attempt to provide a complete specification of every single new feature, but instead provides a brief introduction to each feature. For full details, you should refer to the documentation for Python 2.4, such as the Python Library Reference and the Python Reference Manual. Often you will be referred to the PEP for a particular new feature for explanations of the implementation and design rationale.

PEP 218: Built-In Set Objects

Python 2.3 introduced the sets module. C implementations of set data types have now been added to the Python core as two new built-in types, set(iterable)() and frozenset(iterable)(). They provide high speed operations for membership testing, for eliminating duplicates from sequences, and for mathematical operations like unions, intersections, differences, and symmetric differences.

>>> a = set('abracadabra')              # form a set from a string
>>> 'z' in a                            # fast membership testing
>>> a                                   # unique letters in a
set(['a', 'r', 'b', 'c', 'd'])
>>> ''.join(a)                          # convert back into a string

>>> b = set('alacazam')                 # form a second set
>>> a - b                               # letters in a but not in b
set(['r', 'd', 'b'])
>>> a | b                               # letters in either a or b
set(['a', 'c', 'r', 'd', 'b', 'm', 'z', 'l'])
>>> a & b                               # letters in both a and b
set(['a', 'c'])
>>> a ^ b                               # letters in a or b but not both
set(['r', 'd', 'b', 'm', 'z', 'l'])

>>> a.add('z')                          # add a new element
>>> a.update('wxy')                     # add multiple new elements
>>> a
set(['a', 'c', 'b', 'd', 'r', 'w', 'y', 'x', 'z'])
>>> a.remove('x')                       # take one element out
>>> a
set(['a', 'c', 'b', 'd', 'r', 'w', 'y', 'z'])

The frozenset() type is an immutable version of set(). Since it is immutable and hashable, it may be used as a dictionary key or as a member of another set.

The sets module remains in the standard library, and may be useful if you wish to subclass the Set or ImmutableSet classes. There are currently no plans to deprecate the module.

See also

PEP 218 - Adding a Built-In Set Object Type
Originally proposed by Greg Wilson and ultimately implemented by Raymond Hettinger.

PEP 237: Unifying Long Integers and Integers

The lengthy transition process for this PEP, begun in Python 2.2, takes another step forward in Python 2.4. In 2.3, certain integer operations that would behave differently after int/long unification triggered FutureWarning warnings and returned values limited to 32 or 64 bits (depending on your platform). In 2.4, these expressions no longer produce a warning and instead produce a different result that’s usually a long integer.

The problematic expressions are primarily left shifts and lengthy hexadecimal and octal constants. For example, 2 << 32 results in a warning in 2.3, evaluating to 0 on 32-bit platforms. In Python 2.4, this expression now returns the correct answer, 8589934592.

See also

PEP 237 - Unifying Long Integers and Integers
Original PEP written by Moshe Zadka and GvR. The changes for 2.4 were implemented by Kalle Svensson.

PEP 289: Generator Expressions

The iterator feature introduced in Python 2.2 and the itertools module make it easier to write programs that loop through large data sets without having the entire data set in memory at one time. List comprehensions don’t fit into this picture very well because they produce a Python list object containing all of the items. This unavoidably pulls all of the objects into memory, which can be a problem if your data set is very large. When trying to write a functionally-styled program, it would be natural to write something like:

links = [link for link in get_all_links() if not link.followed]
for link in links:

instead of

for link in get_all_links():
    if link.followed:

The first form is more concise and perhaps more readable, but if you’re dealing with a large number of link objects you’d have to write the second form to avoid having all link objects in memory at the same time.

Generator expressions work similarly to list comprehensions but don’t materialize the entire list; instead they create a generator that will return elements one by one. The above example could be written as:

links = (link for link in get_all_links() if not link.followed)
for link in links:

Generator expressions always have to be written inside parentheses, as in the above example. The parentheses signalling a function call also count, so if you want to create an iterator that will be immediately passed to a function you could write:

print sum(obj.count for obj in list_all_objects())

Generator expressions differ from list comprehensions in various small ways. Most notably, the loop variable (obj in the above example) is not accessible outside of the generator expression. List comprehensions leave the variable assigned to its last value; future versions of Python will change this, making list comprehensions match generator expressions in this respect.

See also

PEP 289 - Generator Expressions
Proposed by Raymond Hettinger and implemented by Jiwon Seo with early efforts steered by Hye-Shik Chang.

PEP 292: Simpler String Substitutions

Some new classes in the standard library provide an alternative mechanism for substituting variables into strings; this style of substitution may be better for applications where untrained users need to edit templates.

The usual way of substituting variables by name is the % operator:

>>> '%(page)i: %(title)s' % {'page':2, 'title': 'The Best of Times'}
'2: The Best of Times'

When writing the template string, it can be easy to forget the i or s after the closing parenthesis. This isn’t a big problem if the template is in a Python module, because you run the code, get an “Unsupported format character” ValueError, and fix the problem. However, consider an application such as Mailman where template strings or translations are being edited by users who aren’t aware of the Python language. The format string’s syntax is complicated to explain to such users, and if they make a mistake, it’s difficult to provide helpful feedback to them.

PEP 292 adds a Template class to the string module that uses $ to indicate a substitution:

>>> import string
>>> t = string.Template('$page: $title')
>>> t.substitute({'page':2, 'title': 'The Best of Times'})
'2: The Best of Times'

If a key is missing from the dictionary, the substitute() method will raise a KeyError. There’s also a safe_substitute() method that ignores missing keys:

>>> t = string.Template('$page: $title')
>>> t.safe_substitute({'page':3})
'3: $title'

See also

PEP 292 - Simpler String Substitutions
Written and implemented by Barry Warsaw.

PEP 318: Decorators for Functions and Methods

Python 2.2 extended Python’s object model by adding static methods and class methods, but it didn’t extend Python’s syntax to provide any new way of defining static or class methods. Instead, you had to write a def statement in the usual way, and pass the resulting method to a staticmethod() or classmethod() function that would wrap up the function as a method of the new type. Your code would look like this:

class C:
   def meth (cls):

   meth = classmethod(meth)   # Rebind name to wrapped-up class method

If the method was very long, it would be easy to miss or forget the classmethod() invocation after the function body.

The intention was always to add some syntax to make such definitions more readable, but at the time of 2.2’s release a good syntax was not obvious. Today a good syntax still isn’t obvious but users are asking for easier access to the feature; a new syntactic feature has been added to meet this need.

The new feature is called “function decorators”. The name comes from the idea that classmethod(), staticmethod(), and friends are storing additional information on a function object; they’re decorating functions with more details.

The notation borrows from Java and uses the '@' character as an indicator. Using the new syntax, the example above would be written:

class C:

   def meth (cls):

The @classmethod is shorthand for the meth=classmethod(meth) assignment. More generally, if you have the following:

def f ():

It’s equivalent to the following pre-decorator code:

def f(): ...
f = A(B(C(f)))

Decorators must come on the line before a function definition, one decorator per line, and can’t be on the same line as the def statement, meaning that @A def f(): ... is illegal. You can only decorate function definitions, either at the module level or inside a class; you can’t decorate class definitions.

A decorator is just a function that takes the function to be decorated as an argument and returns either the same function or some new object. The return value of the decorator need not be callable (though it typically is), unless further decorators will be applied to the result. It’s easy to write your own decorators. The following simple example just sets an attribute on the function object:

>>> def deco(func):
...    func.attr = 'decorated'
...    return func
>>> @deco
... def f(): pass
>>> f
<function f at 0x402ef0d4>
>>> f.attr

As a slightly more realistic example, the following decorator checks that the supplied argument is an integer:

def require_int (func):
    def wrapper (arg):
        assert isinstance(arg, int)
        return func(arg)

    return wrapper

def p1 (arg):
    print arg

def p2(arg):
    print arg*2

An example in PEP 318 contains a fancier version of this idea that lets you both specify the required type and check the returned type.

Decorator functions can take arguments. If arguments are supplied, your decorator function is called with only those arguments and must return a new decorator function; this function must take a single function and return a function, as previously described. In other words, @A @B @C(args) becomes:

def f(): ...
_deco = C(args)
f = A(B(_deco(f)))

Getting this right can be slightly brain-bending, but it’s not too difficult.

A small related change makes the func_name attribute of functions writable. This attribute is used to display function names in tracebacks, so decorators should change the name of any new function that’s constructed and returned.

See also

PEP 318 - Decorators for Functions, Methods and Classes
Written by Kevin D. Smith, Jim Jewett, and Skip Montanaro. Several people wrote patches implementing function decorators, but the one that was actually checked in was patch #979728, written by Mark Russell.
This Wiki page contains several examples of decorators.

PEP 322: Reverse Iteration

A new built-in function, reversed(seq)(), takes a sequence and returns an iterator that loops over the elements of the sequence in reverse order.

>>> for i in reversed(xrange(1,4)):
...    print i

Compared to extended slicing, such as range(1,4)[::-1], reversed() is easier to read, runs faster, and uses substantially less memory.

Note that reversed() only accepts sequences, not arbitrary iterators. If you want to reverse an iterator, first convert it to a list with list().

>>> input = open('/etc/passwd', 'r')
>>> for line in reversed(list(input)):
...   print line
root:*:0:0:System Administrator:/var/root:/bin/tcsh

See also

PEP 322 - Reverse Iteration
Written and implemented by Raymond Hettinger.

PEP 324: New subprocess Module

The standard library provides a number of ways to execute a subprocess, offering different features and different levels of complexity. os.system(command)() is easy to use, but slow (it runs a shell process which executes the command) and dangerous (you have to be careful about escaping the shell’s metacharacters). The popen2 module offers classes that can capture standard output and standard error from the subprocess, but the naming is confusing. The subprocess module cleans this up, providing a unified interface that offers all the features you might need.

Instead of popen2‘s collection of classes, subprocess contains a single class called Popen whose constructor supports a number of different keyword arguments.

class Popen(args, bufsize=0, executable=None,
            stdin=None, stdout=None, stderr=None,
            preexec_fn=None, close_fds=False, shell=False,
            cwd=None, env=None, universal_newlines=False,
            startupinfo=None, creationflags=0):

args is commonly a sequence of strings that will be the arguments to the program executed as the subprocess. (If the shell argument is true, args can be a string which will then be passed on to the shell for interpretation, just as os.system() does.)

stdin, stdout, and stderr specify what the subprocess’s input, output, and error streams will be. You can provide a file object or a file descriptor, or you can use the constant subprocess.PIPE to create a pipe between the subprocess and the parent.

The constructor has a number of handy options:

  • close_fds requests that all file descriptors be closed before running the subprocess.
  • cwd specifies the working directory in which the subprocess will be executed (defaulting to whatever the parent’s working directory is).
  • env is a dictionary specifying environment variables.
  • preexec_fn is a function that gets called before the child is started.
  • universal_newlines opens the child’s input and output using Python’s universal newline feature.

Once you’ve created the Popen instance, you can call its wait() method to pause until the subprocess has exited, poll() to check if it’s exited without pausing, or communicate(data)() to send the string data to the subprocess’s standard input. communicate(data)() then reads any data that the subprocess has sent to its standard output or standard error, returning a tuple (stdout_data, stderr_data).

call() is a shortcut that passes its arguments along to the Popen constructor, waits for the command to complete, and returns the status code of the subprocess. It can serve as a safer analog to os.system():

sts = subprocess.call(['dpkg', '-i', '/tmp/new-package.deb'])
if sts == 0:
    # Success
    # dpkg returned an error

The command is invoked without use of the shell. If you really do want to use the shell, you can add shell=True as a keyword argument and provide a string instead of a sequence:

sts = subprocess.call('dpkg -i /tmp/new-package.deb', shell=True)

The PEP takes various examples of shell and Python code and shows how they’d be translated into Python code that uses subprocess. Reading this section of the PEP is highly recommended.

See also

PEP 324 - subprocess - New process module
Written and implemented by Peter Åstrand, with assistance from Fredrik Lundh and others.

PEP 327: Decimal Data Type

Python has always supported floating-point (FP) numbers, based on the underlying C double type, as a data type. However, while most programming languages provide a floating-point type, many people (even programmers) are unaware that floating-point numbers don’t represent certain decimal fractions accurately. The new Decimal type can represent these fractions accurately, up to a user-specified precision limit.

Why is Decimal needed?

The limitations arise from the representation used for floating-point numbers. FP numbers are made up of three components:

  • The sign, which is positive or negative.
  • The mantissa, which is a single-digit binary number followed by a fractional part. For example, 1.01 in base-2 notation is 1 + 0/2 + 1/4, or 1.25 in decimal notation.
  • The exponent, which tells where the decimal point is located in the number represented.

For example, the number 1.25 has positive sign, a mantissa value of 1.01 (in binary), and an exponent of 0 (the decimal point doesn’t need to be shifted). The number 5 has the same sign and mantissa, but the exponent is 2 because the mantissa is multiplied by 4 (2 to the power of the exponent 2); 1.25 * 4 equals 5.

Modern systems usually provide floating-point support that conforms to a standard called IEEE 754. C’s double type is usually implemented as a 64-bit IEEE 754 number, which uses 52 bits of space for the mantissa. This means that numbers can only be specified to 52 bits of precision. If you’re trying to represent numbers whose expansion repeats endlessly, the expansion is cut off after 52 bits. Unfortunately, most software needs to produce output in base 10, and common fractions in base 10 are often repeating decimals in binary. For example, 1.1 decimal is binary 1.0001100110011 ...; .1 = 1/16 + 1/32 + 1/256 plus an infinite number of additional terms. IEEE 754 has to chop off that infinitely repeated decimal after 52 digits, so the representation is slightly inaccurate.

Sometimes you can see this inaccuracy when the number is printed:

>>> 1.1

The inaccuracy isn’t always visible when you print the number because the FP-to- decimal-string conversion is provided by the C library, and most C libraries try to produce sensible output. Even if it’s not displayed, however, the inaccuracy is still there and subsequent operations can magnify the error.

For many applications this doesn’t matter. If I’m plotting points and displaying them on my monitor, the difference between 1.1 and 1.1000000000000001 is too small to be visible. Reports often limit output to a certain number of decimal places, and if you round the number to two or three or even eight decimal places, the error is never apparent. However, for applications where it does matter, it’s a lot of work to implement your own custom arithmetic routines.

Hence, the Decimal type was created.

The Decimal type

A new module, decimal, was added to Python’s standard library. It contains two classes, Decimal and Context. Decimal instances represent numbers, and Context instances are used to wrap up various settings such as the precision and default rounding mode.

Decimal instances are immutable, like regular Python integers and FP numbers; once it’s been created, you can’t change the value an instance represents. Decimal instances can be created from integers or strings:

>>> import decimal
>>> decimal.Decimal(1972)
>>> decimal.Decimal("1.1")

You can also provide tuples containing the sign, the mantissa represented as a tuple of decimal digits, and the exponent:

>>> decimal.Decimal((1, (1, 4, 7, 5), -2))

Cautionary note: the sign bit is a Boolean value, so 0 is positive and 1 is negative.

Converting from floating-point numbers poses a bit of a problem: should the FP number representing 1.1 turn into the decimal number for exactly 1.1, or for 1.1 plus whatever inaccuracies are introduced? The decision was to dodge the issue and leave such a conversion out of the API. Instead, you should convert the floating-point number into a string using the desired precision and pass the string to the Decimal constructor:

>>> f = 1.1
>>> decimal.Decimal(str(f))
>>> decimal.Decimal('%.12f' % f)

Once you have Decimal instances, you can perform the usual mathematical operations on them. One limitation: exponentiation requires an integer exponent:

>>> a = decimal.Decimal('35.72')
>>> b = decimal.Decimal('1.73')
>>> a+b
>>> a-b
>>> a*b
>>> a/b
>>> a ** 2
>>> a**b
Traceback (most recent call last):
decimal.InvalidOperation: x ** (non-integer)

You can combine Decimal instances with integers, but not with floating- point numbers:

>>> a + 4
>>> a + 4.5
Traceback (most recent call last):
TypeError: You can interact Decimal only with int, long or Decimal data types.

Decimal numbers can be used with the math and cmath modules, but note that they’ll be immediately converted to floating-point numbers before the operation is performed, resulting in a possible loss of precision and accuracy. You’ll also get back a regular floating-point number and not a Decimal.

>>> import math, cmath
>>> d = decimal.Decimal('123456789012.345')
>>> math.sqrt(d)
>>> cmath.sqrt(-d)

Decimal instances have a sqrt() method that returns a Decimal, but if you need other things such as trigonometric functions you’ll have to implement them.

>>> d.sqrt()

The Context type

Instances of the Context class encapsulate several settings for decimal operations:

  • prec is the precision, the number of decimal places.
  • rounding specifies the rounding mode. The decimal module has constants for the various possibilities: ROUND_DOWN, ROUND_CEILING, ROUND_HALF_EVEN, and various others.
  • traps is a dictionary specifying what happens on encountering certain error conditions: either an exception is raised or a value is returned. Some examples of error conditions are division by zero, loss of precision, and overflow.

There’s a thread-local default context available by calling getcontext(); you can change the properties of this context to alter the default precision, rounding, or trap handling. The following example shows the effect of changing the precision of the default context:

>>> decimal.getcontext().prec
>>> decimal.Decimal(1) / decimal.Decimal(7)
>>> decimal.getcontext().prec = 9
>>> decimal.Decimal(1) / decimal.Decimal(7)

The default action for error conditions is selectable; the module can either return a special value such as infinity or not-a-number, or exceptions can be raised:

>>> decimal.Decimal(1) / decimal.Decimal(0)
Traceback (most recent call last):
decimal.DivisionByZero: x / 0
>>> decimal.getcontext().traps[decimal.DivisionByZero] = False
>>> decimal.Decimal(1) / decimal.Decimal(0)

The Context instance also has various methods for formatting numbers such as to_eng_string() and to_sci_string().

For more information, see the documentation for the decimal module, which includes a quick-start tutorial and a reference.

See also

PEP 327 - Decimal Data Type
Written by Facundo Batista and implemented by Facundo Batista, Eric Price, Raymond Hettinger, Aahz, and Tim Peters.
The article uses Fortran code to illustrate many of the problems that floating- point inaccuracy can cause.
A description of a decimal-based representation. This representation is being proposed as a standard, and underlies the new Python decimal type. Much of this material was written by Mike Cowlishaw, designer of the Rexx language.

PEP 328: Multi-line Imports

One language change is a small syntactic tweak aimed at making it easier to import many names from a module. In a from module import names statement, names is a sequence of names separated by commas. If the sequence is very long, you can either write multiple imports from the same module, or you can use backslashes to escape the line endings like this:

from SimpleXMLRPCServer import SimpleXMLRPCServer,\

The syntactic change in Python 2.4 simply allows putting the names within parentheses. Python ignores newlines within a parenthesized expression, so the backslashes are no longer needed:

from SimpleXMLRPCServer import (SimpleXMLRPCServer,

The PEP also proposes that all import statements be absolute imports, with a leading . character to indicate a relative import. This part of the PEP was not implemented for Python 2.4, but was completed for Python 2.5.

See also

PEP 328 - Imports: Multi-Line and Absolute/Relative
Written by Aahz. Multi-line imports were implemented by Dima Dorfman.

PEP 331: Locale-Independent Float/String Conversions

The locale modules lets Python software select various conversions and display conventions that are localized to a particular country or language. However, the module was careful to not change the numeric locale because various functions in Python’s implementation required that the numeric locale remain set to the 'C' locale. Often this was because the code was using the C library’s atof() function.

Not setting the numeric locale caused trouble for extensions that used third- party C libraries, however, because they wouldn’t have the correct locale set. The motivating example was GTK+, whose user interface widgets weren’t displaying numbers in the current locale.

The solution described in the PEP is to add three new functions to the Python API that perform ASCII-only conversions, ignoring the locale setting:

  • PyOS_ascii_strtod(str, ptr)() and PyOS_ascii_atof(str, ptr)() both convert a string to a C double.
  • PyOS_ascii_formatd(buffer, buf_len, format, d)() converts a double to an ASCII string.

The code for these functions came from the GLib library (http://library.gnome.org/devel/glib/stable/), whose developers kindly relicensed the relevant functions and donated them to the Python Software Foundation. The locale module can now change the numeric locale, letting extensions such as GTK+ produce the correct results.

See also

PEP 331 - Locale-Independent Float/String Conversions
Written by Christian R. Reis, and implemented by Gustavo Carneiro.

Other Language Changes

Here are all of the changes that Python 2.4 makes to the core Python language.

  • Decorators for functions and methods were added (PEP 318).

  • Built-in set() and frozenset() types were added (PEP 218). Other new built-ins include the reversed(seq)() function (PEP 322).

  • Generator expressions were added (PEP 289).

  • Certain numeric expressions no longer return values restricted to 32 or 64 bits (PEP 237).

  • You can now put parentheses around the list of names in a from module import names statement (PEP 328).

  • The dict.update() method now accepts the same argument forms as the dict constructor. This includes any mapping, any iterable of key/value pairs, and keyword arguments. (Contributed by Raymond Hettinger.)

  • The string methods ljust(), rjust(), and center() now take an optional argument for specifying a fill character other than a space. (Contributed by Raymond Hettinger.)

  • Strings also gained an rsplit() method that works like the split() method but splits from the end of the string. (Contributed by Sean Reifschneider.)

    >>> 'www.python.org'.split('.', 1)
    ['www', 'python.org']
    'www.python.org'.rsplit('.', 1)
    ['www.python', 'org']
  • Three keyword parameters, cmp, key, and reverse, were added to the sort() method of lists. These parameters make some common usages of sort() simpler. All of these parameters are optional.

    For the cmp parameter, the value should be a comparison function that takes two parameters and returns -1, 0, or +1 depending on how the parameters compare. This function will then be used to sort the list. Previously this was the only parameter that could be provided to sort().

    key should be a single-parameter function that takes a list element and returns a comparison key for the element. The list is then sorted using the comparison keys. The following example sorts a list case-insensitively:

    >>> L = ['A', 'b', 'c', 'D']
    >>> L.sort()                 # Case-sensitive sort
    >>> L
    ['A', 'D', 'b', 'c']
    >>> # Using 'key' parameter to sort list
    >>> L.sort(key=lambda x: x.lower())
    >>> L
    ['A', 'b', 'c', 'D']
    >>> # Old-fashioned way
    >>> L.sort(cmp=lambda x,y: cmp(x.lower(), y.lower()))
    >>> L
    ['A', 'b', 'c', 'D']

    The last example, which uses the cmp parameter, is the old way to perform a case-insensitive sort. It works but is slower than using a key parameter. Using key calls lower() method once for each element in the list while using cmp will call it twice for each comparison, so using key saves on invocations of the lower() method.

    For simple key functions and comparison functions, it is often possible to avoid a lambda expression by using an unbound method instead. For example, the above case-insensitive sort is best written as:

    >>> L.sort(key=str.lower)
    >>> L
    ['A', 'b', 'c', 'D']

    Finally, the reverse parameter takes a Boolean value. If the value is true, the list will be sorted into reverse order. Instead of L.sort() ; L.reverse(), you can now write L.sort(reverse=True).

    The results of sorting are now guaranteed to be stable. This means that two entries with equal keys will be returned in the same order as they were input. For example, you can sort a list of people by name, and then sort the list by age, resulting in a list sorted by age where people with the same age are in name-sorted order.

    (All changes to sort() contributed by Raymond Hettinger.)

  • There is a new built-in function sorted(iterable)() that works like the in-place list.sort() method but can be used in expressions. The differences are:

  • the input may be any iterable;

  • a newly formed copy is sorted, leaving the original intact; and

  • the expression returns the new sorted copy

    >>> L = [9,7,8,3,2,4,1,6,5]
    >>> [10+i for i in sorted(L)]       # usable in a list comprehension
    [11, 12, 13, 14, 15, 16, 17, 18, 19]
    >>> L                               # original is left unchanged
    >>> sorted('Monty Python')          # any iterable may be an input
    [' ', 'M', 'P', 'h', 'n', 'n', 'o', 'o', 't', 't', 'y', 'y']
    >>> # List the contents of a dict sorted by key values
    >>> colormap = dict(red=1, blue=2, green=3, black=4, yellow=5)
    >>> for k, v in sorted(colormap.iteritems()):
    ...     print k, v
    black 4
    blue 2
    green 3
    red 1
    yellow 5

    (Contributed by Raymond Hettinger.)

  • Integer operations will no longer trigger an OverflowWarning. The OverflowWarning warning will disappear in Python 2.5.

  • The interpreter gained a new switch, -m, that takes a name, searches for the corresponding module on sys.path, and runs the module as a script. For example, you can now run the Python profiler with python -m profile. (Contributed by Nick Coghlan.)

  • The eval(expr, globals, locals)() and execfile(filename, globals, locals)() functions and the exec statement now accept any mapping type for the locals parameter. Previously this had to be a regular Python dictionary. (Contributed by Raymond Hettinger.)

  • The zip() built-in function and itertools.izip() now return an empty list if called with no arguments. Previously they raised a TypeError exception. This makes them more suitable for use with variable length argument lists:

    >>> def transpose(array):
    ...    return zip(*array)
    >>> transpose([(1,2,3), (4,5,6)])
    [(1, 4), (2, 5), (3, 6)]
    >>> transpose([])

    (Contributed by Raymond Hettinger.)

  • Encountering a failure while importing a module no longer leaves a partially- initialized module object in sys.modules. The incomplete module object left behind would fool further imports of the same module into succeeding, leading to confusing errors. (Fixed by Tim Peters.)

  • None is now a constant; code that binds a new value to the name None is now a syntax error. (Contributed by Raymond Hettinger.)


  • The inner loops for list and tuple slicing were optimized and now run about one-third faster. The inner loops for dictionaries were also optimized, resulting in performance boosts for keys(), values(), items(), iterkeys(), itervalues(), and iteritems(). (Contributed by Raymond Hettinger.)
  • The machinery for growing and shrinking lists was optimized for speed and for space efficiency. Appending and popping from lists now runs faster due to more efficient code paths and less frequent use of the underlying system realloc(). List comprehensions also benefit. list.extend() was also optimized and no longer converts its argument into a temporary list before extending the base list. (Contributed by Raymond Hettinger.)
  • list(), tuple(), map(), filter(), and zip() now run several times faster with non-sequence arguments that supply a __len__() method. (Contributed by Raymond Hettinger.)
  • The methods list.__getitem__(), dict.__getitem__(), and dict.__contains__() are are now implemented as method_descriptor objects rather than wrapper_descriptor objects. This form of access doubles their performance and makes them more suitable for use as arguments to functionals: map(mydict.__getitem__, keylist). (Contributed by Raymond Hettinger.)
  • Added a new opcode, LIST_APPEND, that simplifies the generated bytecode for list comprehensions and speeds them up by about a third. (Contributed by Raymond Hettinger.)
  • The peephole bytecode optimizer has been improved to produce shorter, faster bytecode; remarkably, the resulting bytecode is more readable. (Enhanced by Raymond Hettinger.)
  • String concatenations in statements of the form s = s + "abc" and s += "abc" are now performed more efficiently in certain circumstances. This optimization won’t be present in other Python implementations such as Jython, so you shouldn’t rely on it; using the join() method of strings is still recommended when you want to efficiently glue a large number of strings together. (Contributed by Armin Rigo.)

The net result of the 2.4 optimizations is that Python 2.4 runs the pystone benchmark around 5% faster than Python 2.3 and 35% faster than Python 2.2. (pystone is not a particularly good benchmark, but it’s the most commonly used measurement of Python’s performance. Your own applications may show greater or smaller benefits from Python 2.4.)

New, Improved, and Deprecated Modules

As usual, Python’s standard library received a number of enhancements and bug fixes. Here’s a partial list of the most notable changes, sorted alphabetically by module name. Consult the Misc/NEWS file in the source tree for a more complete list of changes, or look through the CVS logs for all the details.

  • The asyncore module’s loop() function now has a count parameter that lets you perform a limited number of passes through the polling loop. The default is still to loop forever.

  • The base64 module now has more complete RFC 3548 support for Base64, Base32, and Base16 encoding and decoding, including optional case folding and optional alternative alphabets. (Contributed by Barry Warsaw.)

  • The bisect module now has an underlying C implementation for improved performance. (Contributed by Dmitry Vasiliev.)

  • The CJKCodecs collections of East Asian codecs, maintained by Hye-Shik Chang, was integrated into 2.4. The new encodings are:

  • Chinese (PRC): gb2312, gbk, gb18030, big5hkscs, hz

  • Chinese (ROC): big5, cp950

  • Japanese: cp932, euc-jis-2004, euc-jp, euc-jisx0213, iso-2022-jp,

    iso-2022-jp-1, iso-2022-jp-2, iso-2022-jp-3, iso-2022-jp-ext, iso-2022-jp-2004, shift-jis, shift-jisx0213, shift-jis-2004

  • Korean: cp949, euc-kr, johab, iso-2022-kr

  • Some other new encodings were added: HP Roman8, ISO_8859-11, ISO_8859-16, PCTP-154, and TIS-620.

  • The UTF-8 and UTF-16 codecs now cope better with receiving partial input. Previously the StreamReader class would try to read more data, making it impossible to resume decoding from the stream. The read() method will now return as much data as it can and future calls will resume decoding where previous ones left off. (Implemented by Walter Dörwald.)

  • There is a new collections module for various specialized collection datatypes. Currently it contains just one type, deque, a double- ended queue that supports efficiently adding and removing elements from either end:

    >>> from collections import deque
    >>> d = deque('ghi')        # make a new deque with three items
    >>> d.append('j')           # add a new entry to the right side
    >>> d.appendleft('f')       # add a new entry to the left side
    >>> d                       # show the representation of the deque
    deque(['f', 'g', 'h', 'i', 'j'])
    >>> d.pop()                 # return and remove the rightmost item
    >>> d.popleft()             # return and remove the leftmost item
    >>> list(d)                 # list the contents of the deque
    ['g', 'h', 'i']
    >>> 'h' in d                # search the deque

    Several modules, such as the Queue and threading modules, now take advantage of collections.deque for improved performance. (Contributed by Raymond Hettinger.)

  • The ConfigParser classes have been enhanced slightly. The read() method now returns a list of the files that were successfully parsed, and the set() method raises TypeError if passed a value argument that isn’t a string. (Contributed by John Belmonte and David Goodger.)

  • The curses module now supports the ncurses extension use_default_colors(). On platforms where the terminal supports transparency, this makes it possible to use a transparent background. (Contributed by Jörg Lehmann.)

  • The difflib module now includes an HtmlDiff class that creates an HTML table showing a side by side comparison of two versions of a text. (Contributed by Dan Gass.)

  • The email package was updated to version 3.0, which dropped various deprecated APIs and removes support for Python versions earlier than 2.3. The 3.0 version of the package uses a new incremental parser for MIME messages, available in the email.FeedParser module. The new parser doesn’t require reading the entire message into memory, and doesn’t throw exceptions if a message is malformed; instead it records any problems in the defect attribute of the message. (Developed by Anthony Baxter, Barry Warsaw, Thomas Wouters, and others.)

  • The heapq module has been converted to C. The resulting tenfold improvement in speed makes the module suitable for handling high volumes of data. In addition, the module has two new functions nlargest() and nsmallest() that use heaps to find the N largest or smallest values in a dataset without the expense of a full sort. (Contributed by Raymond Hettinger.)

  • The httplib module now contains constants for HTTP status codes defined in various HTTP-related RFC documents. Constants have names such as OK, CREATED, CONTINUE, and MOVED_PERMANENTLY; use pydoc to get a full list. (Contributed by Andrew Eland.)

  • The imaplib module now supports IMAP’s THREAD command (contributed by Yves Dionne) and new deleteacl() and myrights() methods (contributed by Arnaud Mazin).

  • The itertools module gained a groupby(iterable[, *func*])() function. iterable is something that can be iterated over to return a stream of elements, and the optional func parameter is a function that takes an element and returns a key value; if omitted, the key is simply the element itself. groupby() then groups the elements into subsequences which have matching values of the key, and returns a series of 2-tuples containing the key value and an iterator over the subsequence.

    Here’s an example to make this clearer. The key function simply returns whether a number is even or odd, so the result of groupby() is to return consecutive runs of odd or even numbers.

    >>> import itertools
    >>> L = [2, 4, 6, 7, 8, 9, 11, 12, 14]
    >>> for key_val, it in itertools.groupby(L, lambda x: x % 2):
    ...    print key_val, list(it)
    0 [2, 4, 6]
    1 [7]
    0 [8]
    1 [9, 11]
    0 [12, 14]

    groupby() is typically used with sorted input. The logic for groupby() is similar to the Unix uniq filter which makes it handy for eliminating, counting, or identifying duplicate elements:

    >>> word = 'abracadabra'
    >>> letters = sorted(word)   # Turn string into a sorted list of letters
    >>> letters
    ['a', 'a', 'a', 'a', 'a', 'b', 'b', 'c', 'd', 'r', 'r']
    >>> for k, g in itertools.groupby(letters):
    ...    print k, list(g)
    a ['a', 'a', 'a', 'a', 'a']
    b ['b', 'b']
    c ['c']
    d ['d']
    r ['r', 'r']
    >>> # List unique letters
    >>> [k for k, g in groupby(letters)]
    ['a', 'b', 'c', 'd', 'r']
    >>> # Count letter occurrences
    >>> [(k, len(list(g))) for k, g in groupby(letters)]
    [('a', 5), ('b', 2), ('c', 1), ('d', 1), ('r', 2)]

    (Contributed by Hye-Shik Chang.)

  • itertools also gained a function named tee(iterator, N)() that returns N independent iterators that replicate iterator. If N is omitted, the default is 2.

    >>> L = [1,2,3]
    >>> i1, i2 = itertools.tee(L)
    >>> i1,i2
    (<itertools.tee object at 0x402c2080>, <itertools.tee object at 0x402c2090>)
    >>> list(i1)               # Run the first iterator to exhaustion
    [1, 2, 3]
    >>> list(i2)               # Run the second iterator to exhaustion
    [1, 2, 3]

    Note that tee() has to keep copies of the values returned by the iterator; in the worst case, it may need to keep all of them. This should therefore be used carefully if the leading iterator can run far ahead of the trailing iterator in a long stream of inputs. If the separation is large, then you might as well use list() instead. When the iterators track closely with one another, tee() is ideal. Possible applications include bookmarking, windowing, or lookahead iterators. (Contributed by Raymond Hettinger.)

  • A number of functions were added to the locale module, such as bind_textdomain_codeset() to specify a particular encoding and a family of l*gettext() functions that return messages in the chosen encoding. (Contributed by Gustavo Niemeyer.)

  • Some keyword arguments were added to the logging package’s basicConfig() function to simplify log configuration. The default behavior is to log messages to standard error, but various keyword arguments can be specified to log to a particular file, change the logging format, or set the logging level. For example:

    import logging
        level=0,  # Log all messages

    Other additions to the logging package include a log(level, msg)() convenience method, as well as a TimedRotatingFileHandler class that rotates its log files at a timed interval. The module already had RotatingFileHandler, which rotated logs once the file exceeded a certain size. Both classes derive from a new BaseRotatingHandler class that can be used to implement other rotating handlers.

    (Changes implemented by Vinay Sajip.)

  • The marshal module now shares interned strings on unpacking a data structure. This may shrink the size of certain pickle strings, but the primary effect is to make .pyc files significantly smaller. (Contributed by Martin von Löwis.)

  • The nntplib module’s NNTP class gained description() and descriptions() methods to retrieve newsgroup descriptions for a single group or for a range of groups. (Contributed by Jürgen A. Erhard.)

  • Two new functions were added to the operator module, attrgetter(attr)() and itemgetter(index)(). Both functions return callables that take a single argument and return the corresponding attribute or item; these callables make excellent data extractors when used with map() or sorted(). For example:

    >>> L = [('c', 2), ('d', 1), ('a', 4), ('b', 3)]
    >>> map(operator.itemgetter(0), L)
    ['c', 'd', 'a', 'b']
    >>> map(operator.itemgetter(1), L)
    [2, 1, 4, 3]
    >>> sorted(L, key=operator.itemgetter(1)) # Sort list by second tuple item
    [('d', 1), ('c', 2), ('b', 3), ('a', 4)]

    (Contributed by Raymond Hettinger.)

  • The optparse module was updated in various ways. The module now passes its messages through gettext.gettext(), making it possible to internationalize Optik’s help and error messages. Help messages for options can now include the string '%default', which will be replaced by the option’s default value. (Contributed by Greg Ward.)

  • The long-term plan is to deprecate the rfc822 module in some future Python release in favor of the email package. To this end, the email.Utils.formatdate() function has been changed to make it usable as a replacement for rfc822.formatdate(). You may want to write new e-mail processing code with this in mind. (Change implemented by Anthony Baxter.)

  • A new urandom(n)() function was added to the os module, returning a string containing n bytes of random data. This function provides access to platform-specific sources of randomness such as /dev/urandom on Linux or the Windows CryptoAPI. (Contributed by Trevor Perrin.)

  • Another new function: os.path.lexists(path)() returns true if the file specified by path exists, whether or not it’s a symbolic link. This differs from the existing os.path.exists(path)() function, which returns false if path is a symlink that points to a destination that doesn’t exist. (Contributed by Beni Cherniavsky.)

  • A new getsid() function was added to the posix module that underlies the os module. (Contributed by J. Raynor.)

  • The poplib module now supports POP over SSL. (Contributed by Hector Urtubia.)

  • The profile module can now profile C extension functions. (Contributed by Nick Bastin.)

  • The random module has a new method called getrandbits(N)() that returns a long integer N bits in length. The existing randrange() method now uses getrandbits() where appropriate, making generation of arbitrarily large random numbers more efficient. (Contributed by Raymond Hettinger.)

  • The regular expression language accepted by the re module was extended with simple conditional expressions, written as (?(group)A|B). group is either a numeric group ID or a group name defined with (?P<group>...) earlier in the expression. If the specified group matched, the regular expression pattern A will be tested against the string; if the group didn’t match, the pattern B will be used instead. (Contributed by Gustavo Niemeyer.)

  • The re module is also no longer recursive, thanks to a massive amount of work by Gustavo Niemeyer. In a recursive regular expression engine, certain patterns result in a large amount of C stack space being consumed, and it was possible to overflow the stack. For example, if you matched a 30000-byte string of a characters against the expression (a|b)+, one stack frame was consumed per character. Python 2.3 tried to check for stack overflow and raise a RuntimeError exception, but certain patterns could sidestep the checking and if you were unlucky Python could segfault. Python 2.4’s regular expression engine can match this pattern without problems.

  • The signal module now performs tighter error-checking on the parameters to the signal.signal() function. For example, you can’t set a handler on the SIGKILL signal; previous versions of Python would quietly accept this, but 2.4 will raise a RuntimeError exception.

  • Two new functions were added to the socket module. socketpair() returns a pair of connected sockets and getservbyport(port)() looks up the service name for a given port number. (Contributed by Dave Cole and Barry Warsaw.)

  • The sys.exitfunc() function has been deprecated. Code should be using the existing atexit module, which correctly handles calling multiple exit functions. Eventually sys.exitfunc() will become a purely internal interface, accessed only by atexit.

  • The tarfile module now generates GNU-format tar files by default. (Contributed by Lars Gustaebel.)

  • The threading module now has an elegantly simple way to support thread-local data. The module contains a local class whose attribute values are local to different threads.

    import threading
    data = threading.local()
    data.number = 42
    data.url = ('www.python.org', 80)

    Other threads can assign and retrieve their own values for the number and url attributes. You can subclass local to initialize attributes or to add methods. (Contributed by Jim Fulton.)

  • The timeit module now automatically disables periodic garbage collection during the timing loop. This change makes consecutive timings more comparable. (Contributed by Raymond Hettinger.)

  • The weakref module now supports a wider variety of objects including Python functions, class instances, sets, frozensets, deques, arrays, files, sockets, and regular expression pattern objects. (Contributed by Raymond Hettinger.)

  • The xmlrpclib module now supports a multi-call extension for transmitting multiple XML-RPC calls in a single HTTP operation. (Contributed by Brian Quinlan.)

  • The mpz, rotor, and xreadlines modules have been removed.


The cookielib library supports client-side handling for HTTP cookies, mirroring the Cookie module’s server-side cookie support. Cookies are stored in cookie jars; the library transparently stores cookies offered by the web server in the cookie jar, and fetches the cookie from the jar when connecting to the server. As in web browsers, policy objects control whether cookies are accepted or not.

In order to store cookies across sessions, two implementations of cookie jars are provided: one that stores cookies in the Netscape format so applications can use the Mozilla or Lynx cookie files, and one that stores cookies in the same format as the Perl libwww library.

urllib2 has been changed to interact with cookielib: HTTPCookieProcessor manages a cookie jar that is used when accessing URLs.

This module was contributed by John J. Lee.


The doctest module underwent considerable refactoring thanks to Edward Loper and Tim Peters. Testing can still be as simple as running doctest.testmod(), but the refactorings allow customizing the module’s operation in various ways

The new DocTestFinder class extracts the tests from a given object’s docstrings:

def f (x, y):
    """>>> f(2,2)
>>> f(3,2)
    return x*y

finder = doctest.DocTestFinder()

# Get list of DocTest instances
tests = finder.find(f)

The new DocTestRunner class then runs individual tests and can produce a summary of the results:

runner = doctest.DocTestRunner()
for t in tests:
    tried, failed = runner.run(t)


The above example produces the following output:

1 items passed all tests:
   2 tests in f
2 tests in 1 items.
2 passed and 0 failed.
Test passed.

DocTestRunner uses an instance of the OutputChecker class to compare the expected output with the actual output. This class takes a number of different flags that customize its behaviour; ambitious users can also write a completely new subclass of OutputChecker.

The default output checker provides a number of handy features. For example, with the doctest.ELLIPSIS option flag, an ellipsis (...) in the expected output matches any substring, making it easier to accommodate outputs that vary in minor ways:

def o (n):
    """>>> o(1)
<__main__.C instance at 0x...>

Another special string, <BLANKLINE>, matches a blank line:

def p (n):
    """>>> p(1)

Another new capability is producing a diff-style display of the output by specifying the doctest.REPORT_UDIFF (unified diffs), doctest.REPORT_CDIFF (context diffs), or doctest.REPORT_NDIFF (delta-style) option flags. For example:

def g (n):
    """>>> g(4)
    L = 'here is a rather lengthy list of words'.split()
    for word in L[:n]:
        print word

Running the above function’s tests with doctest.REPORT_UDIFF specified, you get the following output:

File "t.py", line 15, in g
Failed example:
Differences (unified diff with -expected +actual):
    @@ -2,3 +2,3 @@

Build and C API Changes

Some of the changes to Python’s build process and to the C API are:

  • Three new convenience macros were added for common return values from extension functions: Py_RETURN_NONE, Py_RETURN_TRUE, and Py_RETURN_FALSE. (Contributed by Brett Cannon.)
  • Another new macro, Py_CLEAR(obj), decreases the reference count of obj and sets obj to the null pointer. (Contributed by Jim Fulton.)
  • A new function, PyTuple_Pack(N, obj1, obj2, ..., objN)(), constructs tuples from a variable length argument list of Python objects. (Contributed by Raymond Hettinger.)
  • A new function, PyDict_Contains(d, k)(), implements fast dictionary lookups without masking exceptions raised during the look-up process. (Contributed by Raymond Hettinger.)
  • The Py_IS_NAN(X) macro returns 1 if its float or double argument X is a NaN. (Contributed by Tim Peters.)
  • C code can avoid unnecessary locking by using the new PyEval_ThreadsInitialized() function to tell if any thread operations have been performed. If this function returns false, no lock operations are needed. (Contributed by Nick Coghlan.)
  • A new function, PyArg_VaParseTupleAndKeywords(), is the same as PyArg_ParseTupleAndKeywords() but takes a va_list instead of a number of arguments. (Contributed by Greg Chapman.)
  • A new method flag, METH_COEXISTS, allows a function defined in slots to co-exist with a PyCFunction having the same name. This can halve the access time for a method such as set.__contains__(). (Contributed by Raymond Hettinger.)
  • Python can now be built with additional profiling for the interpreter itself, intended as an aid to people developing the Python core. Providing ----enable-profiling to the configure script will let you profile the interpreter with gprof, and providing the ----with-tsc switch enables profiling using the Pentium’s Time-Stamp- Counter register. Note that the ----with-tsc switch is slightly misnamed, because the profiling feature also works on the PowerPC platform, though that processor architecture doesn’t call that register “the TSC register”. (Contributed by Jeremy Hylton.)
  • The tracebackobject type has been renamed to PyTracebackObject.

Port-Specific Changes

  • The Windows port now builds under MSVC++ 7.1 as well as version 6. (Contributed by Martin von Löwis.)

Porting to Python 2.4

This section lists previously described changes that may require changes to your code:

  • Left shifts and hexadecimal/octal constants that are too large no longer trigger a FutureWarning and return a value limited to 32 or 64 bits; instead they return a long integer.
  • Integer operations will no longer trigger an OverflowWarning. The OverflowWarning warning will disappear in Python 2.5.
  • The zip() built-in function and itertools.izip() now return an empty list instead of raising a TypeError exception if called with no arguments.
  • You can no longer compare the date and datetime instances provided by the datetime module. Two instances of different classes will now always be unequal, and relative comparisons (<, >) will raise a TypeError.
  • dircache.listdir() now passes exceptions to the caller instead of returning empty lists.
  • LexicalHandler.startDTD() used to receive the public and system IDs in the wrong order. This has been corrected; applications relying on the wrong order need to be fixed.
  • fcntl.ioctl() now warns if the mutate argument is omitted and relevant.
  • The tarfile module now generates GNU-format tar files by default.
  • Encountering a failure while importing a module no longer leaves a partially- initialized module object in sys.modules.
  • None is now a constant; code that binds a new value to the name None is now a syntax error.
  • The signals.signal() function now raises a RuntimeError exception for certain illegal values; previously these errors would pass silently. For example, you can no longer set a handler on the SIGKILL signal.


The author would like to thank the following people for offering suggestions, corrections and assistance with various drafts of this article: Koray Can, Hye- Shik Chang, Michael Dyck, Raymond Hettinger, Brian Hurt, Hamish Lawson, Fredrik Lundh, Sean Reifschneider, Sadruddin Rejeb.