What’s New in Python 2.7
************************

Author:
   A.M. Kuchling (amk at amk.ca)

This article explains the new features in Python 2.7.  Python 2.7 was
released on July 3, 2010.

Numeric handling has been improved in many ways, for both floating-
point numbers and for the "Decimal" class. There are some useful
additions to the standard library, such as a greatly enhanced
"unittest" module, the "argparse" module for parsing command-line
options, convenient "OrderedDict" and "Counter" classes in the
"collections" module, and many other improvements.

Python 2.7 is planned to be the last of the 2.x releases, so we worked
on making it a good release for the long term.  To help with porting
to Python 3, several new features from the Python 3.x series have been
included in 2.7.

This article doesn’t attempt to provide a complete specification of
the new features, but instead provides a convenient overview.  For
full details, you should refer to the documentation for Python 2.7 at
https://docs.python.org. If you want to understand the rationale for
the design and implementation, refer to the PEP for a particular new
feature or the issue on https://bugs.python.org in which a change was
discussed.  Whenever possible, “What’s New in Python” links to the
bug/patch item for each change.


The Future for Python 2.x
=========================

Python 2.7 is the last major release in the 2.x series, as the Python
maintainers have shifted the focus of their new feature development
efforts to the Python 3.x series. This means that while Python 2
continues to receive bug fixes, and to be updated to build correctly
on new hardware and versions of supported operated systems, there will
be no new full feature releases for the language or standard library.

However, while there is a large common subset between Python 2.7 and
Python 3, and many of the changes involved in migrating to that common
subset, or directly to Python 3, can be safely automated, some other
changes (notably those associated with Unicode handling) may require
careful consideration, and preferably robust automated regression test
suites, to migrate effectively.

This means that Python 2.7 will remain in place for a long time,
providing a stable and supported base platform for production systems
that have not yet been ported to Python 3. The full expected lifecycle
of the Python 2.7 series is detailed in **PEP 373**.

Some key consequences of the long-term significance of 2.7 are:

* As noted above, the 2.7 release has a much longer period of
  maintenance when compared to earlier 2.x versions. Python 2.7 is
  currently expected to remain supported by the core development team
  (receiving security updates and other bug fixes) until at least 2020
  (10 years after its initial release, compared to the more typical
  support period of 18–24 months).

* As the Python 2.7 standard library ages, making effective use of the
  Python Package Index (either directly or via a redistributor)
  becomes more important for Python 2 users. In addition to a wide
  variety of third party packages for various tasks, the available
  packages include backports of new modules and features from the
  Python 3 standard library that are compatible with Python 2, as well
  as various tools and libraries that can make it easier to migrate to
  Python 3. The Python Packaging User Guide provides guidance on
  downloading and installing software from the Python Package Index.

* While the preferred approach to enhancing Python 2 is now the
  publication of new packages on the Python Package Index, this
  approach doesn’t necessarily work in all cases, especially those
  related to network security. In exceptional cases that cannot be
  handled adequately by publishing new or updated packages on PyPI,
  the Python Enhancement Proposal process may be used to make the case
  for adding new features directly to the Python 2 standard library.
  Any such additions, and the maintenance releases where they were
  added, will be noted in the New Features Added to Python 2.7
  Maintenance Releases section below.

For projects wishing to migrate from Python 2 to Python 3, or for
library and framework developers wishing to support users on both
Python 2 and Python 3, there are a variety of tools and guides
available to help decide on a suitable approach and manage some of the
technical details involved. The recommended starting point is the
Porting Python 2 Code to Python 3 HOWTO guide.


Changes to the Handling of Deprecation Warnings
===============================================

For Python 2.7, a policy decision was made to silence warnings only of
interest to developers by default.  "DeprecationWarning" and its
descendants are now ignored unless otherwise requested, preventing
users from seeing warnings triggered by an application.  This change
was also made in the branch that became Python 3.2. (Discussed on
stdlib-sig and carried out in bpo-7319.)

In previous releases, "DeprecationWarning" messages were enabled by
default, providing Python developers with a clear indication of where
their code may break in a future major version of Python.

However, there are increasingly many users of Python-based
applications who are not directly involved in the development of those
applications.  "DeprecationWarning" messages are irrelevant to such
users, making them worry about an application that’s actually working
correctly and burdening application developers with responding to
these concerns.

You can re-enable display of "DeprecationWarning" messages by running
Python with the "-Wdefault" (short form: "-Wd") switch, or by setting
the "PYTHONWARNINGS" environment variable to ""default"" (or ""d"")
before running Python.  Python code can also re-enable them by calling
"warnings.simplefilter('default')".

The "unittest" module also automatically reenables deprecation
warnings when running tests.


Python 3.1 Features
===================

Much as Python 2.6 incorporated features from Python 3.0, version 2.7
incorporates some of the new features in Python 3.1.  The 2.x series
continues to provide tools for migrating to the 3.x series.

A partial list of 3.1 features that were backported to 2.7:

* The syntax for set literals ("{1,2,3}" is a mutable set).

* Dictionary and set comprehensions ("{i: i*2 for i in range(3)}").

* Multiple context managers in a single "with" statement.

* A new version of the "io" library, rewritten in C for performance.

* The ordered-dictionary type described in PEP 372: Adding an Ordered
  Dictionary to collections.

* The new "","" format specifier described in PEP 378: Format
  Specifier for Thousands Separator.

* The "memoryview" object.

* A small subset of the "importlib" module, described below.

* The "repr()" of a float "x" is shorter in many cases: it’s now based
  on the shortest decimal string that’s guaranteed to round back to
  "x".  As in previous versions of Python, it’s guaranteed that
  "float(repr(x))" recovers "x".

* Float-to-string and string-to-float conversions are correctly
  rounded. The "round()" function is also now correctly rounded.

* The "PyCapsule" type, used to provide a C API for extension modules.

* The "PyLong_AsLongAndOverflow()" C API function.

Other new Python3-mode warnings include:

* "operator.isCallable()" and "operator.sequenceIncludes()", which are
  not supported in 3.x, now trigger warnings.

* The "-3" switch now automatically enables the "-Qwarn" switch that
  causes warnings about using classic division with integers and long
  integers.


PEP 372: Adding an Ordered Dictionary to collections
====================================================

Regular Python dictionaries iterate over key/value pairs in arbitrary
order. Over the years, a number of authors have written alternative
implementations that remember the order that the keys were originally
inserted.  Based on the experiences from those implementations, 2.7
introduces a new "OrderedDict" class in the "collections" module.

The "OrderedDict" API provides the same interface as regular
dictionaries but iterates over keys and values in a guaranteed order
depending on when a key was first inserted:

   >>> from collections import OrderedDict
   >>> d = OrderedDict([('first', 1),
   ...                  ('second', 2),
   ...                  ('third', 3)])
   >>> d.items()
   [('first', 1), ('second', 2), ('third', 3)]

If a new entry overwrites an existing entry, the original insertion
position is left unchanged:

   >>> d['second'] = 4
   >>> d.items()
   [('first', 1), ('second', 4), ('third', 3)]

Deleting an entry and reinserting it will move it to the end:

   >>> del d['second']
   >>> d['second'] = 5
   >>> d.items()
   [('first', 1), ('third', 3), ('second', 5)]

The "popitem()" method has an optional *last* argument that defaults
to "True".  If *last* is true, the most recently added key is returned
and removed; if it’s false, the oldest key is selected:

   >>> od = OrderedDict([(x,0) for x in range(20)])
   >>> od.popitem()
   (19, 0)
   >>> od.popitem()
   (18, 0)
   >>> od.popitem(last=False)
   (0, 0)
   >>> od.popitem(last=False)
   (1, 0)

Comparing two ordered dictionaries checks both the keys and values,
and requires that the insertion order was the same:

   >>> od1 = OrderedDict([('first', 1),
   ...                    ('second', 2),
   ...                    ('third', 3)])
   >>> od2 = OrderedDict([('third', 3),
   ...                    ('first', 1),
   ...                    ('second', 2)])
   >>> od1 == od2
   False
   >>> # Move 'third' key to the end
   >>> del od2['third']; od2['third'] = 3
   >>> od1 == od2
   True

Comparing an "OrderedDict" with a regular dictionary ignores the
insertion order and just compares the keys and values.

How does the "OrderedDict" work?  It maintains a doubly-linked list of
keys, appending new keys to the list as they’re inserted. A secondary
dictionary maps keys to their corresponding list node, so deletion
doesn’t have to traverse the entire linked list and therefore remains
O(1).

The standard library now supports use of ordered dictionaries in
several modules.

* The "ConfigParser" module uses them by default, meaning that
  configuration files can now be read, modified, and then written back
  in their original order.

* The "_asdict()" method for "collections.namedtuple()" now returns an
  ordered dictionary with the values appearing in the same order as
  the underlying tuple indices.

* The "json" module’s "JSONDecoder" class constructor was extended
  with an *object_pairs_hook* parameter to allow "OrderedDict"
  instances to be built by the decoder. Support was also added for
  third-party tools like PyYAML.

See also:

  **PEP 372** - Adding an ordered dictionary to collections
     PEP written by Armin Ronacher and Raymond Hettinger; implemented
     by Raymond Hettinger.


PEP 378: Format Specifier for Thousands Separator
=================================================

To make program output more readable, it can be useful to add
separators to large numbers, rendering them as
18,446,744,073,709,551,616 instead of 18446744073709551616.

The fully general solution for doing this is the "locale" module,
which can use different separators (“,” in North America, “.” in
Europe) and different grouping sizes, but "locale" is complicated to
use and unsuitable for multi-threaded applications where different
threads are producing output for different locales.

Therefore, a simple comma-grouping mechanism has been added to the
mini-language used by the "str.format()" method.  When formatting a
floating-point number, simply include a comma between the width and
the precision:

   >>> '{:20,.2f}'.format(18446744073709551616.0)
   '18,446,744,073,709,551,616.00'

When formatting an integer, include the comma after the width:

>>> '{:20,d}'.format(18446744073709551616)
'18,446,744,073,709,551,616'

This mechanism is not adaptable at all; commas are always used as the
separator and the grouping is always into three-digit groups.  The
comma-formatting mechanism isn’t as general as the "locale" module,
but it’s easier to use.

See also:

  **PEP 378** - Format Specifier for Thousands Separator
     PEP written by Raymond Hettinger; implemented by Eric Smith.


PEP 389: The argparse Module for Parsing Command Lines
======================================================

The "argparse" module for parsing command-line arguments was added as
a more powerful replacement for the "optparse" module.

This means Python now supports three different modules for parsing
command-line arguments: "getopt", "optparse", and "argparse".  The
"getopt" module closely resembles the C library’s "getopt()" function,
so it remains useful if you’re writing a Python prototype that will
eventually be rewritten in C. "optparse" becomes redundant, but there
are no plans to remove it because there are many scripts still using
it, and there’s no automated way to update these scripts.  (Making the
"argparse" API consistent with "optparse"’s interface was discussed
but rejected as too messy and difficult.)

In short, if you’re writing a new script and don’t need to worry about
compatibility with earlier versions of Python, use "argparse" instead
of "optparse".

Here’s an example:

   import argparse

   parser = argparse.ArgumentParser(description='Command-line example.')

   # Add optional switches
   parser.add_argument('-v', action='store_true', dest='is_verbose',
                       help='produce verbose output')
   parser.add_argument('-o', action='store', dest='output',
                       metavar='FILE',
                       help='direct output to FILE instead of stdout')
   parser.add_argument('-C', action='store', type=int, dest='context',
                       metavar='NUM', default=0,
                       help='display NUM lines of added context')

   # Allow any number of additional arguments.
   parser.add_argument(nargs='*', action='store', dest='inputs',
                       help='input filenames (default is stdin)')

   args = parser.parse_args()
   print args.__dict__

Unless you override it, "-h" and "--help" switches are automatically
added, and produce neatly formatted output:

   -> ./python.exe argparse-example.py --help
   usage: argparse-example.py [-h] [-v] [-o FILE] [-C NUM] [inputs [inputs ...]]

   Command-line example.

   positional arguments:
     inputs      input filenames (default is stdin)

   optional arguments:
     -h, --help  show this help message and exit
     -v          produce verbose output
     -o FILE     direct output to FILE instead of stdout
     -C NUM      display NUM lines of added context

As with "optparse", the command-line switches and arguments are
returned as an object with attributes named by the *dest* parameters:

   -> ./python.exe argparse-example.py -v
   {'output': None,
    'is_verbose': True,
    'context': 0,
    'inputs': []}

   -> ./python.exe argparse-example.py -v -o /tmp/output -C 4 file1 file2
   {'output': '/tmp/output',
    'is_verbose': True,
    'context': 4,
    'inputs': ['file1', 'file2']}

"argparse" has much fancier validation than "optparse"; you can
specify an exact number of arguments as an integer, 0 or more
arguments by passing "'*'", 1 or more by passing "'+'", or an optional
argument with "'?'".  A top-level parser can contain sub-parsers to
define subcommands that have different sets of switches, as in "svn
commit", "svn checkout", etc.  You can specify an argument’s type as
"FileType", which will automatically open files for you and
understands that "'-'" means standard input or output.

See also:

  "argparse" documentation
     The documentation page of the argparse module.

  Upgrading optparse code
     Part of the Python documentation, describing how to convert code
     that uses "optparse".

  **PEP 389** - argparse - New Command Line Parsing Module
     PEP written and implemented by Steven Bethard.


PEP 391: Dictionary-Based Configuration For Logging
===================================================

The "logging" module is very flexible; applications can define a tree
of logging subsystems, and each logger in this tree can filter out
certain messages, format them differently, and direct messages to a
varying number of handlers.

All this flexibility can require a lot of configuration.  You can
write Python statements to create objects and set their properties,
but a complex set-up requires verbose but boring code. "logging" also
supports a "fileConfig()" function that parses a file, but the file
format doesn’t support configuring filters, and it’s messier to
generate programmatically.

Python 2.7 adds a "dictConfig()" function that uses a dictionary to
configure logging.  There are many ways to produce a dictionary from
different sources: construct one with code; parse a file containing
JSON; or use a YAML parsing library if one is installed.  For more
information see Configuration functions.

The following example configures two loggers, the root logger and a
logger named “network”.  Messages sent to the root logger will be sent
to the system log using the syslog protocol, and messages to the
“network” logger will be written to a "network.log" file that will be
rotated once the log reaches 1MB.

   import logging
   import logging.config

   configdict = {
    'version': 1,    # Configuration schema in use; must be 1 for now
    'formatters': {
        'standard': {
            'format': ('%(asctime)s %(name)-15s '
                       '%(levelname)-8s %(message)s')}},

    'handlers': {'netlog': {'backupCount': 10,
                        'class': 'logging.handlers.RotatingFileHandler',
                        'filename': '/logs/network.log',
                        'formatter': 'standard',
                        'level': 'INFO',
                        'maxBytes': 1000000},
                 'syslog': {'class': 'logging.handlers.SysLogHandler',
                            'formatter': 'standard',
                            'level': 'ERROR'}},

    # Specify all the subordinate loggers
    'loggers': {
                'network': {
                            'handlers': ['netlog']
                }
    },
    # Specify properties of the root logger
    'root': {
             'handlers': ['syslog']
    },
   }

   # Set up configuration
   logging.config.dictConfig(configdict)

   # As an example, log two error messages
   logger = logging.getLogger('/')
   logger.error('Database not found')

   netlogger = logging.getLogger('network')
   netlogger.error('Connection failed')

Three smaller enhancements to the "logging" module, all implemented by
Vinay Sajip, are:

* The "SysLogHandler" class now supports syslogging over TCP.  The
  constructor has a *socktype* parameter giving the type of socket to
  use, either "socket.SOCK_DGRAM" for UDP or "socket.SOCK_STREAM" for
  TCP.  The default protocol remains UDP.

* "Logger" instances gained a "getChild()" method that retrieves a
  descendant logger using a relative path. For example, once you
  retrieve a logger by doing "log = getLogger('app')", calling
  "log.getChild('network.listen')" is equivalent to
  "getLogger('app.network.listen')".

* The "LoggerAdapter" class gained an "isEnabledFor()" method that
  takes a *level* and returns whether the underlying logger would
  process a message of that level of importance.

See also:

  **PEP 391** - Dictionary-Based Configuration For Logging
     PEP written and implemented by Vinay Sajip.


PEP 3106: Dictionary Views
==========================

The dictionary methods "keys()", "values()", and "items()" are
different in Python 3.x.  They return an object called a *view*
instead of a fully materialized list.

It’s not possible to change the return values of "keys()", "values()",
and "items()" in Python 2.7 because too much code would break.
Instead the 3.x versions were added under the new names "viewkeys()",
"viewvalues()", and "viewitems()".

   >>> d = dict((i*10, chr(65+i)) for i in range(26))
   >>> d
   {0: 'A', 130: 'N', 10: 'B', 140: 'O', 20: ..., 250: 'Z'}
   >>> d.viewkeys()
   dict_keys([0, 130, 10, 140, 20, 150, 30, ..., 250])

Views can be iterated over, but the key and item views also behave
like sets.  The "&" operator performs intersection, and "|" performs a
union:

   >>> d1 = dict((i*10, chr(65+i)) for i in range(26))
   >>> d2 = dict((i**.5, i) for i in range(1000))
   >>> d1.viewkeys() & d2.viewkeys()
   set([0.0, 10.0, 20.0, 30.0])
   >>> d1.viewkeys() | range(0, 30)
   set([0, 1, 130, 3, 4, 5, 6, ..., 120, 250])

The view keeps track of the dictionary and its contents change as the
dictionary is modified:

   >>> vk = d.viewkeys()
   >>> vk
   dict_keys([0, 130, 10, ..., 250])
   >>> d[260] = '&'
   >>> vk
   dict_keys([0, 130, 260, 10, ..., 250])

However, note that you can’t add or remove keys while you’re iterating
over the view:

   >>> for k in vk:
   ...     d[k*2] = k
   ...
   Traceback (most recent call last):
     File "<stdin>", line 1, in <module>
   RuntimeError: dictionary changed size during iteration

You can use the view methods in Python 2.x code, and the 2to3
converter will change them to the standard "keys()", "values()", and
"items()" methods.

See also:

  **PEP 3106** - Revamping dict.keys(), .values() and .items()
     PEP written by Guido van Rossum. Backported to 2.7 by Alexandre
     Vassalotti; bpo-1967.


PEP 3137: The memoryview Object
===============================

The "memoryview" object provides a view of another object’s memory
content that matches the "bytes" type’s interface.

   >>> import string
   >>> m = memoryview(string.letters)
   >>> m
   <memory at 0x37f850>
   >>> len(m)           # Returns length of underlying object
   52
   >>> m[0], m[25], m[26]   # Indexing returns one byte
   ('a', 'z', 'A')
   >>> m2 = m[0:26]         # Slicing returns another memoryview
   >>> m2
   <memory at 0x37f080>

The content of the view can be converted to a string of bytes or a
list of integers:

   >>> m2.tobytes()
   'abcdefghijklmnopqrstuvwxyz'
   >>> m2.tolist()
   [97, 98, 99, 100, 101, 102, 103, ... 121, 122]
   >>>

"memoryview" objects allow modifying the underlying object if it’s a
mutable object.

   >>> m2[0] = 75
   Traceback (most recent call last):
     File "<stdin>", line 1, in <module>
   TypeError: cannot modify read-only memory
   >>> b = bytearray(string.letters)  # Creating a mutable object
   >>> b
   bytearray(b'abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ')
   >>> mb = memoryview(b)
   >>> mb[0] = '*'         # Assign to view, changing the bytearray.
   >>> b[0:5]              # The bytearray has been changed.
   bytearray(b'*bcde')
   >>>

See also:

  **PEP 3137** - Immutable Bytes and Mutable Buffer
     PEP written by Guido van Rossum. Implemented by Travis Oliphant,
     Antoine Pitrou and others. Backported to 2.7 by Antoine Pitrou;
     bpo-2396.


Other Language Changes
======================

Some smaller changes made to the core Python language are:

* The syntax for set literals has been backported from Python 3.x.
  Curly brackets are used to surround the contents of the resulting
  mutable set; set literals are distinguished from dictionaries by not
  containing colons and values. "{}" continues to represent an empty
  dictionary; use "set()" for an empty set.

     >>> {1, 2, 3, 4, 5}
     set([1, 2, 3, 4, 5])
     >>> set() # empty set
     set([])
     >>> {}    # empty dict
     {}

  Backported by Alexandre Vassalotti; bpo-2335.

* Dictionary and set comprehensions are another feature backported
  from 3.x, generalizing list/generator comprehensions to use the
  literal syntax for sets and dictionaries.

     >>> {x: x*x for x in range(6)}
     {0: 0, 1: 1, 2: 4, 3: 9, 4: 16, 5: 25}
     >>> {('a'*x) for x in range(6)}
     set(['', 'a', 'aa', 'aaa', 'aaaa', 'aaaaa'])

  Backported by Alexandre Vassalotti; bpo-2333.

* The "with" statement can now use multiple context managers in one
  statement.  Context managers are processed from left to right and
  each one is treated as beginning a new "with" statement. This means
  that:

     with A() as a, B() as b:
         ... suite of statements ...

  is equivalent to:

     with A() as a:
         with B() as b:
             ... suite of statements ...

  The "contextlib.nested()" function provides a very similar function,
  so it’s no longer necessary and has been deprecated.

  (Proposed in https://codereview.appspot.com/53094; implemented by
  Georg Brandl.)

* Conversions between floating-point numbers and strings are now
  correctly rounded on most platforms.  These conversions occur in
  many different places: "str()" on floats and complex numbers; the
  "float" and "complex" constructors; numeric formatting; serializing
  and deserializing floats and complex numbers using the "marshal",
  "pickle" and "json" modules; parsing of float and imaginary literals
  in Python code; and "Decimal"-to-float conversion.

  Related to this, the "repr()" of a floating-point number *x* now
  returns a result based on the shortest decimal string that’s
  guaranteed to round back to *x* under correct rounding (with round-
  half-to-even rounding mode).  Previously it gave a string based on
  rounding x to 17 decimal digits.

  The rounding library responsible for this improvement works on
  Windows and on Unix platforms using the gcc, icc, or suncc
  compilers.  There may be a small number of platforms where correct
  operation of this code cannot be guaranteed, so the code is not used
  on such systems.  You can find out which code is being used by
  checking "sys.float_repr_style",  which will be "short" if the new
  code is in use and "legacy" if it isn’t.

  Implemented by Eric Smith and Mark Dickinson, using David Gay’s
  "dtoa.c" library; bpo-7117.

* Conversions from long integers and regular integers to floating
  point now round differently, returning the floating-point number
  closest to the number.  This doesn’t matter for small integers that
  can be converted exactly, but for large numbers that will
  unavoidably lose precision, Python 2.7 now approximates more
  closely.  For example, Python 2.6 computed the following:

     >>> n = 295147905179352891391
     >>> float(n)
     2.9514790517935283e+20
     >>> n - long(float(n))
     65535L

  Python 2.7’s floating-point result is larger, but much closer to the
  true value:

     >>> n = 295147905179352891391
     >>> float(n)
     2.9514790517935289e+20
     >>> n - long(float(n))
     -1L

  (Implemented by Mark Dickinson; bpo-3166.)

  Integer division is also more accurate in its rounding behaviours.
  (Also implemented by Mark Dickinson; bpo-1811.)

* Implicit coercion for complex numbers has been removed; the
  interpreter will no longer ever attempt to call a "__coerce__()"
  method on complex objects.  (Removed by Meador Inge and Mark
  Dickinson; bpo-5211.)

* The "str.format()" method now supports automatic numbering of the
  replacement fields.  This makes using "str.format()" more closely
  resemble using "%s" formatting:

     >>> '{}:{}:{}'.format(2009, 04, 'Sunday')
     '2009:4:Sunday'
     >>> '{}:{}:{day}'.format(2009, 4, day='Sunday')
     '2009:4:Sunday'

  The auto-numbering takes the fields from left to right, so the first
  "{...}" specifier will use the first argument to "str.format()", the
  next specifier will use the next argument, and so on.  You can’t mix
  auto-numbering and explicit numbering – either number all of your
  specifier fields or none of them – but you can mix auto-numbering
  and named fields, as in the second example above.  (Contributed by
  Eric Smith; bpo-5237.)

  Complex numbers now correctly support usage with "format()", and
  default to being right-aligned. Specifying a precision or comma-
  separation applies to both the real and imaginary parts of the
  number, but a specified field width and alignment is applied to the
  whole of the resulting "1.5+3j" output.  (Contributed by Eric Smith;
  bpo-1588 and bpo-7988.)

  The ‘F’ format code now always formats its output using uppercase
  characters, so it will now produce ‘INF’ and ‘NAN’. (Contributed by
  Eric Smith; bpo-3382.)

  A low-level change: the "object.__format__()" method now triggers a
  "PendingDeprecationWarning" if it’s passed a format string, because
  the "__format__()" method for "object" converts the object to a
  string representation and formats that.  Previously the method
  silently applied the format string to the string representation, but
  that could hide mistakes in Python code.  If you’re supplying
  formatting information such as an alignment or precision, presumably
  you’re expecting the formatting to be applied in some object-
  specific way.  (Fixed by Eric Smith; bpo-7994.)

* The "int()" and "long()" types gained a "bit_length" method that
  returns the number of bits necessary to represent its argument in
  binary:

     >>> n = 37
     >>> bin(n)
     '0b100101'
     >>> n.bit_length()
     6
     >>> n = 2**123-1
     >>> n.bit_length()
     123
     >>> (n+1).bit_length()
     124

  (Contributed by Fredrik Johansson and Victor Stinner; bpo-3439.)

* The "import" statement will no longer try an absolute import if a
  relative import (e.g. "from .os import sep") fails.  This fixes a
  bug, but could possibly break certain "import" statements that were
  only working by accident.  (Fixed by Meador Inge; bpo-7902.)

* It’s now possible for a subclass of the built-in "unicode" type to
  override the "__unicode__()" method.  (Implemented by Victor
  Stinner; bpo-1583863.)

* The "bytearray" type’s "translate()" method now accepts "None" as
  its first argument.  (Fixed by Georg Brandl; bpo-4759.)

* When using "@classmethod" and "@staticmethod" to wrap methods as
  class or static methods, the wrapper object now exposes the wrapped
  function as their "__func__" attribute. (Contributed by Amaury
  Forgeot d’Arc, after a suggestion by George Sakkis; bpo-5982.)

* When a restricted set of attributes were set using "__slots__",
  deleting an unset attribute would not raise "AttributeError" as you
  would expect.  Fixed by Benjamin Peterson; bpo-7604.)

* Two new encodings are now supported: “cp720”, used primarily for
  Arabic text; and “cp858”, a variant of CP 850 that adds the euro
  symbol.  (CP720 contributed by Alexander Belchenko and Amaury
  Forgeot d’Arc in bpo-1616979; CP858 contributed by Tim Hatch in
  bpo-8016.)

* The "file" object will now set the "filename" attribute on the
  "IOError" exception when trying to open a directory on POSIX
  platforms (noted by Jan Kaliszewski; bpo-4764), and now explicitly
  checks for and forbids writing to read-only file objects instead of
  trusting the C library to catch and report the error (fixed by
  Stefan Krah; bpo-5677).

* The Python tokenizer now translates line endings itself, so the
  "compile()" built-in function now accepts code using any line-ending
  convention.  Additionally, it no longer requires that the code end
  in a newline.

* Extra parentheses in function definitions are illegal in Python 3.x,
  meaning that you get a syntax error from "def f((x)): pass".  In
  Python3-warning mode, Python 2.7 will now warn about this odd usage.
  (Noted by James Lingard; bpo-7362.)

* It’s now possible to create weak references to old-style class
  objects.  New-style classes were always weak-referenceable.  (Fixed
  by Antoine Pitrou; bpo-8268.)

* When a module object is garbage-collected, the module’s dictionary
  is now only cleared if no one else is holding a reference to the
  dictionary (bpo-7140).


Interpreter Changes
-------------------

A new environment variable, "PYTHONWARNINGS", allows controlling
warnings.  It should be set to a string containing warning settings,
equivalent to those used with the "-W" switch, separated by commas.
(Contributed by Brian Curtin; bpo-7301.)

For example, the following setting will print warnings every time they
occur, but turn warnings from the "Cookie" module into an error.  (The
exact syntax for setting an environment variable varies across
operating systems and shells.)

   export PYTHONWARNINGS=all,error:::Cookie:0


Optimizations
-------------

Several performance enhancements have been added:

* A new opcode was added to perform the initial setup for "with"
  statements, looking up the "__enter__()" and "__exit__()" methods.
  (Contributed by Benjamin Peterson.)

* The garbage collector now performs better for one common usage
  pattern: when many objects are being allocated without deallocating
  any of them.  This would previously take quadratic time for garbage
  collection, but now the number of full garbage collections is
  reduced as the number of objects on the heap grows. The new logic
  only performs a full garbage collection pass when the middle
  generation has been collected 10 times and when the number of
  survivor objects from the middle generation exceeds 10% of the
  number of objects in the oldest generation.  (Suggested by Martin
  von Löwis and implemented by Antoine Pitrou; bpo-4074.)

* The garbage collector tries to avoid tracking simple containers
  which can’t be part of a cycle. In Python 2.7, this is now true for
  tuples and dicts containing atomic types (such as ints, strings,
  etc.). Transitively, a dict containing tuples of atomic types won’t
  be tracked either. This helps reduce the cost of each garbage
  collection by decreasing the number of objects to be considered and
  traversed by the collector. (Contributed by Antoine Pitrou;
  bpo-4688.)

* Long integers are now stored internally either in base "2**15" or in
  base "2**30", the base being determined at build time.  Previously,
  they were always stored in base "2**15".  Using base "2**30" gives
  significant performance improvements on 64-bit machines, but
  benchmark results on 32-bit machines have been mixed.  Therefore,
  the default is to use base "2**30" on 64-bit machines and base
  "2**15" on 32-bit machines; on Unix, there’s a new configure option
  "--enable-big-digits" that can be used to override this default.

  Apart from the performance improvements this change should be
  invisible to end users, with one exception: for testing and
  debugging purposes there’s a new structseq "sys.long_info" that
  provides information about the internal format, giving the number of
  bits per digit and the size in bytes of the C type used to store
  each digit:

     >>> import sys
     >>> sys.long_info
     sys.long_info(bits_per_digit=30, sizeof_digit=4)

  (Contributed by Mark Dickinson; bpo-4258.)

  Another set of changes made long objects a few bytes smaller: 2
  bytes smaller on 32-bit systems and 6 bytes on 64-bit. (Contributed
  by Mark Dickinson; bpo-5260.)

* The division algorithm for long integers has been made faster by
  tightening the inner loop, doing shifts instead of multiplications,
  and fixing an unnecessary extra iteration. Various benchmarks show
  speedups of between 50% and 150% for long integer divisions and
  modulo operations. (Contributed by Mark Dickinson; bpo-5512.)
  Bitwise operations are also significantly faster (initial patch by
  Gregory Smith; bpo-1087418).

* The implementation of "%" checks for the left-side operand being a
  Python string and special-cases it; this results in a 1–3%
  performance increase for applications that frequently use "%" with
  strings, such as templating libraries. (Implemented by Collin
  Winter; bpo-5176.)

* List comprehensions with an "if" condition are compiled into faster
  bytecode.  (Patch by Antoine Pitrou, back-ported to 2.7 by Jeffrey
  Yasskin; bpo-4715.)

* Converting an integer or long integer to a decimal string was made
  faster by special-casing base 10 instead of using a generalized
  conversion function that supports arbitrary bases. (Patch by Gawain
  Bolton; bpo-6713.)

* The "split()", "replace()", "rindex()", "rpartition()", and
  "rsplit()" methods of string-like types (strings, Unicode strings,
  and "bytearray" objects) now use a fast reverse-search algorithm
  instead of a character-by-character scan.  This is sometimes faster
  by a factor of 10.  (Added by Florent Xicluna; bpo-7462 and
  bpo-7622.)

* The "pickle" and "cPickle" modules now automatically intern the
  strings used for attribute names, reducing memory usage of the
  objects resulting from unpickling.  (Contributed by Jake McGuire;
  bpo-5084.)

* The "cPickle" module now special-cases dictionaries, nearly halving
  the time required to pickle them. (Contributed by Collin Winter;
  bpo-5670.)


New and Improved Modules
========================

As in every release, 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 Subversion logs for all the details.

* The "bdb" module’s base debugging class "Bdb" gained a feature for
  skipping modules.  The constructor now takes an iterable containing
  glob-style patterns such as "django.*"; the debugger will not step
  into stack frames from a module that matches one of these patterns.
  (Contributed by Maru Newby after a suggestion by Senthil Kumaran;
  bpo-5142.)

* The "binascii" module now supports the buffer API, so it can be used
  with "memoryview" instances and other similar buffer objects.
  (Backported from 3.x by Florent Xicluna; bpo-7703.)

* Updated module: the "bsddb" module has been updated from 4.7.2devel9
  to version 4.8.4 of the pybsddb package. The new version features
  better Python 3.x compatibility, various bug fixes, and adds several
  new BerkeleyDB flags and methods. (Updated by Jesús Cea Avión;
  bpo-8156.  The pybsddb changelog can be read at
  http://hg.jcea.es/pybsddb/file/tip/ChangeLog.)

* The "bz2" module’s "BZ2File" now supports the context management
  protocol, so you can write "with bz2.BZ2File(...) as f:".
  (Contributed by Hagen Fürstenau; bpo-3860.)

* New class: the "Counter" class in the "collections" module is useful
  for tallying data.  "Counter" instances behave mostly like
  dictionaries but return zero for missing keys instead of raising a
  "KeyError":

     >>> from collections import Counter
     >>> c = Counter()
     >>> for letter in 'here is a sample of english text':
     ...   c[letter] += 1
     ...
     >>> c 
     Counter({' ': 6, 'e': 5, 's': 3, 'a': 2, 'i': 2, 'h': 2,
     'l': 2, 't': 2, 'g': 1, 'f': 1, 'm': 1, 'o': 1, 'n': 1,
     'p': 1, 'r': 1, 'x': 1})
     >>> c['e']
     5
     >>> c['z']
     0

  There are three additional "Counter" methods. "most_common()"
  returns the N most common elements and their counts.  "elements()"
  returns an iterator over the contained elements, repeating each
  element as many times as its count. "subtract()" takes an iterable
  and subtracts one for each element instead of adding; if the
  argument is a dictionary or another "Counter", the counts are
  subtracted.

     >>> c.most_common(5)
     [(' ', 6), ('e', 5), ('s', 3), ('a', 2), ('i', 2)]
     >>> c.elements() ->
        'a', 'a', ' ', ' ', ' ', ' ', ' ', ' ',
        'e', 'e', 'e', 'e', 'e', 'g', 'f', 'i', 'i',
        'h', 'h', 'm', 'l', 'l', 'o', 'n', 'p', 's',
        's', 's', 'r', 't', 't', 'x'
     >>> c['e']
     5
     >>> c.subtract('very heavy on the letter e')
     >>> c['e']    # Count is now lower
     -1

  Contributed by Raymond Hettinger; bpo-1696199.

  New class: "OrderedDict" is described in the earlier section PEP
  372: Adding an Ordered Dictionary to collections.

  New method: The "deque" data type now has a "count()" method that
  returns the number of contained elements equal to the supplied
  argument *x*, and a "reverse()" method that reverses the elements of
  the deque in-place.  "deque" also exposes its maximum length as the
  read-only "maxlen" attribute. (Both features added by Raymond
  Hettinger.)

  The "namedtuple" class now has an optional *rename* parameter. If
  *rename* is true, field names that are invalid because they’ve been
  repeated or aren’t legal Python identifiers will be renamed to legal
  names that are derived from the field’s position within the list of
  fields:

  >>> from collections import namedtuple
  >>> T = namedtuple('T', ['field1', '$illegal', 'for', 'field2'], rename=True)
  >>> T._fields
  ('field1', '_1', '_2', 'field2')

  (Added by Raymond Hettinger; bpo-1818.)

  Finally, the "Mapping" abstract base class now returns
  "NotImplemented" if a mapping is compared to another type that isn’t
  a "Mapping". (Fixed by Daniel Stutzbach; bpo-8729.)

* Constructors for the parsing classes in the "ConfigParser" module
  now take an *allow_no_value* parameter, defaulting to false; if
  true, options without values will be allowed.  For example:

     >>> import ConfigParser, StringIO
     >>> sample_config = """
     ... [mysqld]
     ... user = mysql
     ... pid-file = /var/run/mysqld/mysqld.pid
     ... skip-bdb
     ... """
     >>> config = ConfigParser.RawConfigParser(allow_no_value=True)
     >>> config.readfp(StringIO.StringIO(sample_config))
     >>> config.get('mysqld', 'user')
     'mysql'
     >>> print config.get('mysqld', 'skip-bdb')
     None
     >>> print config.get('mysqld', 'unknown')
     Traceback (most recent call last):
       ...
     NoOptionError: No option 'unknown' in section: 'mysqld'

  (Contributed by Mats Kindahl; bpo-7005.)

* Deprecated function: "contextlib.nested()", which allows handling
  more than one context manager with a single "with" statement, has
  been deprecated, because the "with" statement now supports multiple
  context managers.

* The "cookielib" module now ignores cookies that have an invalid
  version field, one that doesn’t contain an integer value.  (Fixed by
  John J. Lee; bpo-3924.)

* The "copy" module’s "deepcopy()" function will now correctly copy
  bound instance methods.  (Implemented by Robert Collins; bpo-1515.)

* The "ctypes" module now always converts "None" to a C "NULL" pointer
  for arguments declared as pointers.  (Changed by Thomas Heller;
  bpo-4606.)  The underlying libffi library has been updated to
  version 3.0.9, containing various fixes for different platforms.
  (Updated by Matthias Klose; bpo-8142.)

* New method: the "datetime" module’s "timedelta" class gained a
  "total_seconds()" method that returns the number of seconds in the
  duration.  (Contributed by Brian Quinlan; bpo-5788.)

* New method: the "Decimal" class gained a "from_float()" class method
  that performs an exact conversion of a floating-point number to a
  "Decimal". This exact conversion strives for the closest decimal
  approximation to the floating-point representation’s value; the
  resulting decimal value will therefore still include the inaccuracy,
  if any. For example, "Decimal.from_float(0.1)" returns "Decimal('0.
  1000000000000000055511151231257827021181583404541015625')".
  (Implemented by Raymond Hettinger; bpo-4796.)

  Comparing instances of "Decimal" with floating-point numbers now
  produces sensible results based on the numeric values of the
  operands.  Previously such comparisons would fall back to Python’s
  default rules for comparing objects, which produced arbitrary
  results based on their type.  Note that you still cannot combine
  "Decimal" and floating-point in other operations such as addition,
  since you should be explicitly choosing how to convert between float
  and "Decimal".  (Fixed by Mark Dickinson; bpo-2531.)

  The constructor for "Decimal" now accepts floating-point numbers
  (added by Raymond Hettinger; bpo-8257) and non-European Unicode
  characters such as Arabic-Indic digits (contributed by Mark
  Dickinson; bpo-6595).

  Most of the methods of the "Context" class now accept integers as
  well as "Decimal" instances; the only exceptions are the
  "canonical()" and "is_canonical()" methods.  (Patch by Juan José
  Conti; bpo-7633.)

  When using "Decimal" instances with a string’s "format()" method,
  the default alignment was previously left-alignment.  This has been
  changed to right-alignment, which is more sensible for numeric
  types.  (Changed by Mark Dickinson; bpo-6857.)

  Comparisons involving a signaling NaN value (or "sNAN") now signal
  "InvalidOperation" instead of silently returning a true or false
  value depending on the comparison operator.  Quiet NaN values (or
  "NaN") are now hashable.  (Fixed by Mark Dickinson; bpo-7279.)

* The "difflib" module now produces output that is more compatible
  with modern **diff**/**patch** tools through one small change, using
  a tab character instead of spaces as a separator in the header
  giving the filename.  (Fixed by Anatoly Techtonik; bpo-7585.)

* The Distutils "sdist" command now always regenerates the "MANIFEST"
  file, since even if the "MANIFEST.in" or "setup.py" files haven’t
  been modified, the user might have created some new files that
  should be included. (Fixed by Tarek Ziadé; bpo-8688.)

* The "doctest" module’s "IGNORE_EXCEPTION_DETAIL" flag will now
  ignore the name of the module containing the exception being tested.
  (Patch by Lennart Regebro; bpo-7490.)

* The "email" module’s "Message" class will now accept a Unicode-
  valued payload, automatically converting the payload to the encoding
  specified by "output_charset". (Added by R. David Murray;
  bpo-1368247.)

* The "Fraction" class now accepts a single float or "Decimal"
  instance, or two rational numbers, as arguments to its constructor.
  (Implemented by Mark Dickinson; rationals added in bpo-5812, and
  float/decimal in bpo-8294.)

  Ordering comparisons ("<", "<=", ">", ">=") between fractions and
  complex numbers now raise a "TypeError". This fixes an oversight,
  making the "Fraction" match the other numeric types.

* New class: "FTP_TLS" in the "ftplib" module provides secure FTP
  connections using TLS encapsulation of authentication as well as
  subsequent control and data transfers. (Contributed by Giampaolo
  Rodola; bpo-2054.)

  The "storbinary()" method for binary uploads can now restart uploads
  thanks to an added *rest* parameter (patch by Pablo Mouzo;
  bpo-6845.)

* New class decorator: "total_ordering()" in the "functools" module
  takes a class that defines an "__eq__()" method and one of
  "__lt__()", "__le__()", "__gt__()", or "__ge__()", and generates the
  missing comparison methods.  Since the "__cmp__()" method is being
  deprecated in Python 3.x, this decorator makes it easier to define
  ordered classes. (Added by Raymond Hettinger; bpo-5479.)

  New function: "cmp_to_key()" will take an old-style comparison
  function that expects two arguments and return a new callable that
  can be used as the *key* parameter to functions such as "sorted()",
  "min()" and "max()", etc.  The primary intended use is to help with
  making code compatible with Python 3.x. (Added by Raymond
  Hettinger.)

* New function: the "gc" module’s "is_tracked()" returns true if a
  given instance is tracked by the garbage collector, false otherwise.
  (Contributed by Antoine Pitrou; bpo-4688.)

* The "gzip" module’s "GzipFile" now supports the context management
  protocol, so you can write "with gzip.GzipFile(...) as f:"
  (contributed by Hagen Fürstenau; bpo-3860), and it now implements
  the "io.BufferedIOBase" ABC, so you can wrap it with
  "io.BufferedReader" for faster processing (contributed by Nir Aides;
  bpo-7471). It’s also now possible to override the modification time
  recorded in a gzipped file by providing an optional timestamp to the
  constructor.  (Contributed by Jacques Frechet; bpo-4272.)

  Files in gzip format can be padded with trailing zero bytes; the
  "gzip" module will now consume these trailing bytes.  (Fixed by
  Tadek Pietraszek and Brian Curtin; bpo-2846.)

* New attribute: the "hashlib" module now has an "algorithms"
  attribute containing a tuple naming the supported algorithms. In
  Python 2.7, "hashlib.algorithms" contains "('md5', 'sha1', 'sha224',
  'sha256', 'sha384', 'sha512')". (Contributed by Carl Chenet;
  bpo-7418.)

* The default "HTTPResponse" class used by the "httplib" module now
  supports buffering, resulting in much faster reading of HTTP
  responses. (Contributed by Kristján Valur Jónsson; bpo-4879.)

  The "HTTPConnection" and "HTTPSConnection" classes now support a
  *source_address* parameter, a "(host, port)" 2-tuple giving the
  source address that will be used for the connection. (Contributed by
  Eldon Ziegler; bpo-3972.)

* The "ihooks" module now supports relative imports.  Note that
  "ihooks" is an older module for customizing imports, superseded by
  the "imputil" module added in Python 2.0. (Relative import support
  added by Neil Schemenauer.)

* The "imaplib" module now supports IPv6 addresses. (Contributed by
  Derek Morr; bpo-1655.)

* New function: the "inspect" module’s "getcallargs()" takes a
  callable and its positional and keyword arguments, and figures out
  which of the callable’s parameters will receive each argument,
  returning a dictionary mapping argument names to their values.  For
  example:

     >>> from inspect import getcallargs
     >>> def f(a, b=1, *pos, **named):
     ...     pass
     >>> getcallargs(f, 1, 2, 3)
     {'a': 1, 'b': 2, 'pos': (3,), 'named': {}}
     >>> getcallargs(f, a=2, x=4)
     {'a': 2, 'b': 1, 'pos': (), 'named': {'x': 4}}
     >>> getcallargs(f)
     Traceback (most recent call last):
     ...
     TypeError: f() takes at least 1 argument (0 given)

  Contributed by George Sakkis; bpo-3135.

* Updated module: The "io" library has been upgraded to the version
  shipped with Python 3.1.  For 3.1, the I/O library was entirely
  rewritten in C and is 2 to 20 times faster depending on the task
  being performed.  The original Python version was renamed to the
  "_pyio" module.

  One minor resulting change: the "io.TextIOBase" class now has an
  "errors" attribute giving the error setting used for encoding and
  decoding errors (one of "'strict'", "'replace'", "'ignore'").

  The "io.FileIO" class now raises an "OSError" when passed an invalid
  file descriptor.  (Implemented by Benjamin Peterson; bpo-4991.)  The
  "truncate()" method now preserves the file position; previously it
  would change the file position to the end of the new file.  (Fixed
  by Pascal Chambon; bpo-6939.)

* New function: "itertools.compress(data, selectors)" takes two
  iterators.  Elements of *data* are returned if the corresponding
  value in *selectors* is true:

     itertools.compress('ABCDEF', [1,0,1,0,1,1]) =>
       A, C, E, F

  New function: "itertools.combinations_with_replacement(iter, r)"
  returns all the possible *r*-length combinations of elements from
  the iterable *iter*.  Unlike "combinations()", individual elements
  can be repeated in the generated combinations:

     itertools.combinations_with_replacement('abc', 2) =>
       ('a', 'a'), ('a', 'b'), ('a', 'c'),
       ('b', 'b'), ('b', 'c'), ('c', 'c')

  Note that elements are treated as unique depending on their position
  in the input, not their actual values.

  The "itertools.count()" function now has a *step* argument that
  allows incrementing by values other than 1.  "count()" also now
  allows keyword arguments, and using non-integer values such as
  floats or "Decimal" instances.  (Implemented by Raymond Hettinger;
  bpo-5032.)

  "itertools.combinations()" and "itertools.product()" previously
  raised "ValueError" for values of *r* larger than the input
  iterable.  This was deemed a specification error, so they now return
  an empty iterator.  (Fixed by Raymond Hettinger; bpo-4816.)

* Updated module: The "json" module was upgraded to version 2.0.9 of
  the simplejson package, which includes a C extension that makes
  encoding and decoding faster. (Contributed by Bob Ippolito;
  bpo-4136.)

  To support the new "collections.OrderedDict" type, "json.load()" now
  has an optional *object_pairs_hook* parameter that will be called
  with any object literal that decodes to a list of pairs.
  (Contributed by Raymond Hettinger; bpo-5381.)

* The "mailbox" module’s "Maildir" class now records the timestamp on
  the directories it reads, and only re-reads them if the modification
  time has subsequently changed.  This improves performance by
  avoiding unneeded directory scans.  (Fixed by A.M. Kuchling and
  Antoine Pitrou; bpo-1607951, bpo-6896.)

* New functions: the "math" module gained "erf()" and "erfc()" for the
  error function and the complementary error function, "expm1()" which
  computes "e**x - 1" with more precision than using "exp()" and
  subtracting 1, "gamma()" for the Gamma function, and "lgamma()" for
  the natural log of the Gamma function. (Contributed by Mark
  Dickinson and nirinA raseliarison; bpo-3366.)

* The "multiprocessing" module’s "Manager*" classes can now be passed
  a callable that will be called whenever a subprocess is started,
  along with a set of arguments that will be passed to the callable.
  (Contributed by lekma; bpo-5585.)

  The "Pool" class, which controls a pool of worker processes, now has
  an optional *maxtasksperchild* parameter.  Worker processes will
  perform the specified number of tasks and then exit, causing the
  "Pool" to start a new worker.  This is useful if tasks may leak
  memory or other resources, or if some tasks will cause the worker to
  become very large. (Contributed by Charles Cazabon; bpo-6963.)

* The "nntplib" module now supports IPv6 addresses. (Contributed by
  Derek Morr; bpo-1664.)

* New functions: the "os" module wraps the following POSIX system
  calls: "getresgid()" and "getresuid()", which return the real,
  effective, and saved GIDs and UIDs; "setresgid()" and "setresuid()",
  which set real, effective, and saved GIDs and UIDs to new values;
  "initgroups()", which initialize the group access list for the
  current process.  (GID/UID functions contributed by Travis H.;
  bpo-6508.  Support for initgroups added by Jean-Paul Calderone;
  bpo-7333.)

  The "os.fork()" function now re-initializes the import lock in the
  child process; this fixes problems on Solaris when "fork()" is
  called from a thread.  (Fixed by Zsolt Cserna; bpo-7242.)

* In the "os.path" module, the "normpath()" and "abspath()" functions
  now preserve Unicode; if their input path is a Unicode string, the
  return value is also a Unicode string. ("normpath()" fixed by Matt
  Giuca in bpo-5827; "abspath()" fixed by Ezio Melotti in bpo-3426.)

* The "pydoc" module now has help for the various symbols that Python
  uses.  You can now do "help('<<')" or "help('@')", for example.
  (Contributed by David Laban; bpo-4739.)

* The "re" module’s "split()", "sub()", and "subn()" now accept an
  optional *flags* argument, for consistency with the other functions
  in the module.  (Added by Gregory P. Smith.)

* New function: "run_path()" in the "runpy" module will execute the
  code at a provided *path* argument.  *path* can be the path of a
  Python source file ("example.py"), a compiled bytecode file
  ("example.pyc"), a directory ("./package/"), or a zip archive
  ("example.zip").  If a directory or zip path is provided, it will be
  added to the front of "sys.path" and the module "__main__" will be
  imported.  It’s expected that the directory or zip contains a
  "__main__.py"; if it doesn’t, some other "__main__.py" might be
  imported from a location later in "sys.path".  This makes more of
  the machinery of "runpy" available to scripts that want to mimic the
  way Python’s command line processes an explicit path name. (Added by
  Nick Coghlan; bpo-6816.)

* New function: in the "shutil" module, "make_archive()" takes a
  filename, archive type (zip or tar-format), and a directory path,
  and creates an archive containing the directory’s contents. (Added
  by Tarek Ziadé.)

  "shutil"’s "copyfile()" and "copytree()" functions now raise a
  "SpecialFileError" exception when asked to copy a named pipe.
  Previously the code would treat named pipes like a regular file by
  opening them for reading, and this would block indefinitely.  (Fixed
  by Antoine Pitrou; bpo-3002.)

* The "signal" module no longer re-installs the signal handler unless
  this is truly necessary, which fixes a bug that could make it
  impossible to catch the EINTR signal robustly.  (Fixed by Charles-
  Francois Natali; bpo-8354.)

* New functions: in the "site" module, three new functions return
  various site- and user-specific paths. "getsitepackages()" returns a
  list containing all global site-packages directories,
  "getusersitepackages()" returns the path of the user’s site-packages
  directory, and "getuserbase()" returns the value of the "USER_BASE"
  environment variable, giving the path to a directory that can be
  used to store data. (Contributed by Tarek Ziadé; bpo-6693.)

  The "site" module now reports exceptions occurring when the
  "sitecustomize" module is imported, and will no longer catch and
  swallow the "KeyboardInterrupt" exception.  (Fixed by Victor
  Stinner; bpo-3137.)

* The "create_connection()" function gained a *source_address*
  parameter, a "(host, port)" 2-tuple giving the source address that
  will be used for the connection. (Contributed by Eldon Ziegler;
  bpo-3972.)

  The "recv_into()" and "recvfrom_into()" methods will now write into
  objects that support the buffer API, most usefully the "bytearray"
  and "memoryview" objects.  (Implemented by Antoine Pitrou;
  bpo-8104.)

* The "SocketServer" module’s "TCPServer" class now supports socket
  timeouts and disabling the Nagle algorithm. The
  "disable_nagle_algorithm" class attribute defaults to "False"; if
  overridden to be true, new request connections will have the
  TCP_NODELAY option set to prevent buffering many small sends into a
  single TCP packet. The "timeout" class attribute can hold a timeout
  in seconds that will be applied to the request socket; if no request
  is received within that time, "handle_timeout()" will be called and
  "handle_request()" will return. (Contributed by Kristján Valur
  Jónsson; bpo-6192 and bpo-6267.)

* Updated module: the "sqlite3" module has been updated to version
  2.6.0 of the pysqlite package. Version 2.6.0 includes a number of
  bugfixes, and adds the ability to load SQLite extensions from shared
  libraries. Call the "enable_load_extension(True)" method to enable
  extensions, and then call "load_extension()" to load a particular
  shared library. (Updated by Gerhard Häring.)

* The "ssl" module’s "SSLSocket" objects now support the buffer API,
  which fixed a test suite failure (fix by Antoine Pitrou; bpo-7133)
  and automatically set OpenSSL’s "SSL_MODE_AUTO_RETRY", which will
  prevent an error code being returned from "recv()" operations that
  trigger an SSL renegotiation (fix by Antoine Pitrou; bpo-8222).

  The "ssl.wrap_socket()" constructor function now takes a *ciphers*
  argument that’s a string listing the encryption algorithms to be
  allowed; the format of the string is described in the OpenSSL
  documentation. (Added by Antoine Pitrou; bpo-8322.)

  Another change makes the extension load all of OpenSSL’s ciphers and
  digest algorithms so that they’re all available.  Some SSL
  certificates couldn’t be verified, reporting an “unknown algorithm”
  error.  (Reported by Beda Kosata, and fixed by Antoine Pitrou;
  bpo-8484.)

  The version of OpenSSL being used is now available as the module
  attributes "ssl.OPENSSL_VERSION" (a string),
  "ssl.OPENSSL_VERSION_INFO" (a 5-tuple), and
  "ssl.OPENSSL_VERSION_NUMBER" (an integer).  (Added by Antoine
  Pitrou; bpo-8321.)

* The "struct" module will no longer silently ignore overflow errors
  when a value is too large for a particular integer format code (one
  of "bBhHiIlLqQ"); it now always raises a "struct.error" exception.
  (Changed by Mark Dickinson; bpo-1523.)  The "pack()" function will
  also attempt to use "__index__()" to convert and pack non-integers
  before trying the "__int__()" method or reporting an error. (Changed
  by Mark Dickinson; bpo-8300.)

* New function: the "subprocess" module’s "check_output()" runs a
  command with a specified set of arguments and returns the command’s
  output as a string when the command runs without error, or raises a
  "CalledProcessError" exception otherwise.

     >>> subprocess.check_output(['df', '-h', '.'])
     'Filesystem     Size   Used  Avail Capacity  Mounted on\n
     /dev/disk0s2    52G    49G   3.0G    94%    /\n'

     >>> subprocess.check_output(['df', '-h', '/bogus'])
       ...
     subprocess.CalledProcessError: Command '['df', '-h', '/bogus']' returned non-zero exit status 1

  (Contributed by Gregory P. Smith.)

  The "subprocess" module will now retry its internal system calls on
  receiving an "EINTR" signal.  (Reported by several people; final
  patch by Gregory P. Smith in bpo-1068268.)

* New function: "is_declared_global()" in the "symtable" module
  returns true for variables that are explicitly declared to be
  global, false for ones that are implicitly global. (Contributed by
  Jeremy Hylton.)

* The "syslog" module will now use the value of "sys.argv[0]" as the
  identifier instead of the previous default value of "'python'".
  (Changed by Sean Reifschneider; bpo-8451.)

* The "sys.version_info" value is now a named tuple, with attributes
  named "major", "minor", "micro", "releaselevel", and "serial".
  (Contributed by Ross Light; bpo-4285.)

  "sys.getwindowsversion()" also returns a named tuple, with
  attributes named "major", "minor", "build", "platform",
  "service_pack", "service_pack_major", "service_pack_minor",
  "suite_mask", and "product_type".  (Contributed by Brian Curtin;
  bpo-7766.)

* The "tarfile" module’s default error handling has changed, to no
  longer suppress fatal errors.  The default error level was
  previously 0, which meant that errors would only result in a message
  being written to the debug log, but because the debug log is not
  activated by default, these errors go unnoticed.  The default error
  level is now 1, which raises an exception if there’s an error.
  (Changed by Lars Gustäbel; bpo-7357.)

  "tarfile" now supports filtering the "TarInfo" objects being added
  to a tar file.  When you call "add()", you may supply an optional
  *filter* argument that’s a callable.  The *filter* callable will be
  passed the "TarInfo" for every file being added, and can modify and
  return it. If the callable returns "None", the file will be excluded
  from the resulting archive.  This is more powerful than the existing
  *exclude* argument, which has therefore been deprecated. (Added by
  Lars Gustäbel; bpo-6856.) The "TarFile" class also now supports the
  context management protocol. (Added by Lars Gustäbel; bpo-7232.)

* The "wait()" method of the "threading.Event" class now returns the
  internal flag on exit.  This means the method will usually return
  true because "wait()" is supposed to block until the internal flag
  becomes true.  The return value will only be false if a timeout was
  provided and the operation timed out. (Contributed by Tim Lesher;
  bpo-1674032.)

* The Unicode database provided by the "unicodedata" module is now
  used internally to determine which characters are numeric,
  whitespace, or represent line breaks.  The database also includes
  information from the "Unihan.txt" data file (patch by Anders
  Chrigström and Amaury Forgeot d’Arc; bpo-1571184) and has been
  updated to version 5.2.0 (updated by Florent Xicluna; bpo-8024).

* The "urlparse" module’s "urlsplit()" now handles unknown URL schemes
  in a fashion compliant with **RFC 3986**: if the URL is of the form
  ""<something>://..."", the text before the "://" is treated as the
  scheme, even if it’s a made-up scheme that the module doesn’t know
  about.  This change may break code that worked around the old
  behaviour.  For example, Python 2.6.4 or 2.5 will return the
  following:

     >>> import urlparse
     >>> urlparse.urlsplit('invented://host/filename?query')
     ('invented', '', '//host/filename?query', '', '')

  Python 2.7 (and Python 2.6.5) will return:

     >>> import urlparse
     >>> urlparse.urlsplit('invented://host/filename?query')
     ('invented', 'host', '/filename?query', '', '')

  (Python 2.7 actually produces slightly different output, since it
  returns a named tuple instead of a standard tuple.)

  The "urlparse" module also supports IPv6 literal addresses as
  defined by **RFC 2732** (contributed by Senthil Kumaran; bpo-2987).

     >>> urlparse.urlparse('http://[1080::8:800:200C:417A]/foo')
     ParseResult(scheme='http', netloc='[1080::8:800:200C:417A]',
                 path='/foo', params='', query='', fragment='')

* New class: the "WeakSet" class in the "weakref" module is a set that
  only holds weak references to its elements; elements will be removed
  once there are no references pointing to them. (Originally
  implemented in Python 3.x by Raymond Hettinger, and backported to
  2.7 by Michael Foord.)

* The ElementTree library, "xml.etree", no longer escapes ampersands
  and angle brackets when outputting an XML processing instruction
  (which looks like "<?xml-stylesheet href="#style1"?>") or comment
  (which looks like "<!-- comment -->"). (Patch by Neil Muller;
  bpo-2746.)

* The XML-RPC client and server, provided by the "xmlrpclib" and
  "SimpleXMLRPCServer" modules, have improved performance by
  supporting HTTP/1.1 keep-alive and by optionally using gzip encoding
  to compress the XML being exchanged.  The gzip compression is
  controlled by the "encode_threshold" attribute of
  "SimpleXMLRPCRequestHandler", which contains a size in bytes;
  responses larger than this will be compressed. (Contributed by
  Kristján Valur Jónsson; bpo-6267.)

* The "zipfile" module’s "ZipFile" now supports the context management
  protocol, so you can write "with zipfile.ZipFile(...) as f:".
  (Contributed by Brian Curtin; bpo-5511.)

  "zipfile" now also supports archiving empty directories and extracts
  them correctly.  (Fixed by Kuba Wieczorek; bpo-4710.) Reading files
  out of an archive is faster, and interleaving "read()" and
  "readline()" now works correctly. (Contributed by Nir Aides;
  bpo-7610.)

  The "is_zipfile()" function now accepts a file object, in addition
  to the path names accepted in earlier versions.  (Contributed by
  Gabriel Genellina; bpo-4756.)

  The "writestr()" method now has an optional *compress_type*
  parameter that lets you override the default compression method
  specified in the "ZipFile" constructor.  (Contributed by Ronald
  Oussoren; bpo-6003.)


New module: importlib
---------------------

Python 3.1 includes the "importlib" package, a re-implementation of
the logic underlying Python’s "import" statement. "importlib" is
useful for implementors of Python interpreters and to users who wish
to write new importers that can participate in the import process.
Python 2.7 doesn’t contain the complete "importlib" package, but
instead has a tiny subset that contains a single function,
"import_module()".

"import_module(name, package=None)" imports a module.  *name* is a
string containing the module or package’s name.  It’s possible to do
relative imports by providing a string that begins with a "."
character, such as "..utils.errors".  For relative imports, the
*package* argument must be provided and is the name of the package
that will be used as the anchor for the relative import.
"import_module()" both inserts the imported module into "sys.modules"
and returns the module object.

Here are some examples:

   >>> from importlib import import_module
   >>> anydbm = import_module('anydbm')  # Standard absolute import
   >>> anydbm
   <module 'anydbm' from '/p/python/Lib/anydbm.py'>
   >>> # Relative import
   >>> file_util = import_module('..file_util', 'distutils.command')
   >>> file_util
   <module 'distutils.file_util' from '/python/Lib/distutils/file_util.pyc'>

"importlib" was implemented by Brett Cannon and introduced in Python
3.1.


New module: sysconfig
---------------------

The "sysconfig" module has been pulled out of the Distutils package,
becoming a new top-level module in its own right. "sysconfig" provides
functions for getting information about Python’s build process:
compiler switches, installation paths, the platform name, and whether
Python is running from its source directory.

Some of the functions in the module are:

* "get_config_var()" returns variables from Python’s Makefile and the
  "pyconfig.h" file.

* "get_config_vars()" returns a dictionary containing all of the
  configuration variables.

* "get_path()" returns the configured path for a particular type of
  module: the standard library, site-specific modules, platform-
  specific modules, etc.

* "is_python_build()" returns true if you’re running a binary from a
  Python source tree, and false otherwise.

Consult the "sysconfig" documentation for more details and for a
complete list of functions.

The Distutils package and "sysconfig" are now maintained by Tarek
Ziadé, who has also started a Distutils2 package (source repository at
https://hg.python.org/distutils2/) for developing a next-generation
version of Distutils.


ttk: Themed Widgets for Tk
--------------------------

Tcl/Tk 8.5 includes a set of themed widgets that re-implement basic Tk
widgets but have a more customizable appearance and can therefore more
closely resemble the native platform’s widgets.  This widget set was
originally called Tile, but was renamed to Ttk (for “themed Tk”) on
being added to Tcl/Tck release 8.5.

To learn more, read the "ttk" module documentation.  You may also wish
to read the Tcl/Tk manual page describing the Ttk theme engine,
available at https://www.tcl.tk/man/tcl8.5/TkCmd/ttk_intro.htm. Some
screenshots of the Python/Ttk code in use are at
https://code.google.com/archive/p/python-ttk/wikis/Screenshots.wiki.

The "ttk" module was written by Guilherme Polo and added in bpo-2983.
An alternate version called "Tile.py", written by Martin Franklin and
maintained by Kevin Walzer, was proposed for inclusion in bpo-2618,
but the authors argued that Guilherme Polo’s work was more
comprehensive.


Updated module: unittest
------------------------

The "unittest" module was greatly enhanced; many new features were
added.  Most of these features were implemented by Michael Foord,
unless otherwise noted.  The enhanced version of the module is
downloadable separately for use with Python versions 2.4 to 2.6,
packaged as the "unittest2" package, from
https://pypi.org/project/unittest2.

When used from the command line, the module can automatically discover
tests.  It’s not as fancy as py.test or nose, but provides a simple
way to run tests kept within a set of package directories.  For
example, the following command will search the "test/" subdirectory
for any importable test files named "test*.py":

   python -m unittest discover -s test

Consult the "unittest" module documentation for more details.
(Developed in bpo-6001.)

The "main()" function supports some other new options:

* "-b" or "--buffer" will buffer the standard output and standard
  error streams during each test.  If the test passes, any resulting
  output will be discarded; on failure, the buffered output will be
  displayed.

* "-c" or "--catch" will cause the control-C interrupt to be handled
  more gracefully.  Instead of interrupting the test process
  immediately, the currently running test will be completed and then
  the partial results up to the interruption will be reported. If
  you’re impatient, a second press of control-C will cause an
  immediate interruption.

  This control-C handler tries to avoid causing problems when the code
  being tested or the tests being run have defined a signal handler of
  their own, by noticing that a signal handler was already set and
  calling it.  If this doesn’t work for you, there’s a
  "removeHandler()" decorator that can be used to mark tests that
  should have the control-C handling disabled.

* "-f" or "--failfast" makes test execution stop immediately when a
  test fails instead of continuing to execute further tests.
  (Suggested by Cliff Dyer and implemented by Michael Foord;
  bpo-8074.)

The progress messages now show ‘x’ for expected failures and ‘u’ for
unexpected successes when run in verbose mode. (Contributed by
Benjamin Peterson.)

Test cases can raise the "SkipTest" exception to skip a test
(bpo-1034053).

The error messages for "assertEqual()", "assertTrue()", and
"assertFalse()" failures now provide more information.  If you set the
"longMessage" attribute of your "TestCase" classes to true, both the
standard error message and any additional message you provide will be
printed for failures.  (Added by Michael Foord; bpo-5663.)

The "assertRaises()" method now returns a context handler when called
without providing a callable object to run.  For example, you can
write this:

   with self.assertRaises(KeyError):
       {}['foo']

(Implemented by Antoine Pitrou; bpo-4444.)

Module- and class-level setup and teardown fixtures are now supported.
Modules can contain "setUpModule()" and "tearDownModule()" functions.
Classes can have "setUpClass()" and "tearDownClass()" methods that
must be defined as class methods (using "@classmethod" or equivalent).
These functions and methods are invoked when the test runner switches
to a test case in a different module or class.

The methods "addCleanup()" and "doCleanups()" were added.
"addCleanup()" lets you add cleanup functions that will be called
unconditionally (after "setUp()" if "setUp()" fails, otherwise after
"tearDown()"). This allows for much simpler resource allocation and
deallocation during tests (bpo-5679).

A number of new methods were added that provide more specialized
tests.  Many of these methods were written by Google engineers for use
in their test suites; Gregory P. Smith, Michael Foord, and GvR worked
on merging them into Python’s version of "unittest".

* "assertIsNone()" and "assertIsNotNone()" take one expression and
  verify that the result is or is not "None".

* "assertIs()" and "assertIsNot()" take two values and check whether
  the two values evaluate to the same object or not. (Added by Michael
  Foord; bpo-2578.)

* "assertIsInstance()" and "assertNotIsInstance()" check whether the
  resulting object is an instance of a particular class, or of one of
  a tuple of classes.  (Added by Georg Brandl; bpo-7031.)

* "assertGreater()", "assertGreaterEqual()", "assertLess()", and
  "assertLessEqual()" compare two quantities.

* "assertMultiLineEqual()" compares two strings, and if they’re not
  equal, displays a helpful comparison that highlights the differences
  in the two strings.  This comparison is now used by default when
  Unicode strings are compared with "assertEqual()".

* "assertRegexpMatches()" and "assertNotRegexpMatches()" checks
  whether the first argument is a string matching or not matching the
  regular expression provided as the second argument (bpo-8038).

* "assertRaisesRegexp()" checks whether a particular exception is
  raised, and then also checks that the string representation of the
  exception matches the provided regular expression.

* "assertIn()" and "assertNotIn()" tests whether *first* is or is not
  in  *second*.

* "assertItemsEqual()" tests whether two provided sequences contain
  the same elements.

* "assertSetEqual()" compares whether two sets are equal, and only
  reports the differences between the sets in case of error.

* Similarly, "assertListEqual()" and "assertTupleEqual()" compare the
  specified types and explain any differences without necessarily
  printing their full values; these methods are now used by default
  when comparing lists and tuples using "assertEqual()". More
  generally, "assertSequenceEqual()" compares two sequences and can
  optionally check whether both sequences are of a particular type.

* "assertDictEqual()" compares two dictionaries and reports the
  differences; it’s now used by default when you compare two
  dictionaries using "assertEqual()".  "assertDictContainsSubset()"
  checks whether all of the key/value pairs in *first* are found in
  *second*.

* "assertAlmostEqual()" and "assertNotAlmostEqual()" test whether
  *first* and *second* are approximately equal.  This method can
  either round their difference to an optionally-specified number of
  *places* (the default is 7) and compare it to zero, or require the
  difference to be smaller than a supplied *delta* value.

* "loadTestsFromName()" properly honors the "suiteClass" attribute of
  the "TestLoader". (Fixed by Mark Roddy; bpo-6866.)

* A new hook lets you extend the "assertEqual()" method to handle new
  data types.  The "addTypeEqualityFunc()" method takes a type object
  and a function. The function will be used when both of the objects
  being compared are of the specified type.  This function should
  compare the two objects and raise an exception if they don’t match;
  it’s a good idea for the function to provide additional information
  about why the two objects aren’t matching, much as the new sequence
  comparison methods do.

"unittest.main()" now takes an optional "exit" argument.  If false,
"main()" doesn’t call "sys.exit()", allowing "main()" to be used from
the interactive interpreter. (Contributed by J. Pablo Fernández;
bpo-3379.)

"TestResult" has new "startTestRun()" and "stopTestRun()" methods that
are called immediately before and after a test run.  (Contributed by
Robert Collins; bpo-5728.)

With all these changes, the "unittest.py" was becoming awkwardly
large, so the module was turned into a package and the code split into
several files (by Benjamin Peterson).  This doesn’t affect how the
module is imported or used.

See also:

  http://www.voidspace.org.uk/python/articles/unittest2.shtml
     Describes the new features, how to use them, and the rationale
     for various design decisions.  (By Michael Foord.)


Updated module: ElementTree 1.3
-------------------------------

The version of the ElementTree library included with Python was
updated to version 1.3.  Some of the new features are:

* The various parsing functions now take a *parser* keyword argument
  giving an "XMLParser" instance that will be used.  This makes it
  possible to override the file’s internal encoding:

     p = ET.XMLParser(encoding='utf-8')
     t = ET.XML("""<root/>""", parser=p)

  Errors in parsing XML now raise a "ParseError" exception, whose
  instances have a "position" attribute containing a (*line*,
  *column*) tuple giving the location of the problem.

* ElementTree’s code for converting trees to a string has been
  significantly reworked, making it roughly twice as fast in many
  cases.  The "ElementTree.write()" and "Element.write()" methods now
  have a *method* parameter that can be “xml” (the default), “html”,
  or “text”.  HTML mode will output empty elements as
  "<empty></empty>" instead of "<empty/>", and text mode will skip
  over elements and only output the text chunks.  If you set the "tag"
  attribute of an element to "None" but leave its children in place,
  the element will be omitted when the tree is written out, so you
  don’t need to do more extensive rearrangement to remove a single
  element.

  Namespace handling has also been improved.  All "xmlns:<whatever>"
  declarations are now output on the root element, not scattered
  throughout the resulting XML.  You can set the default namespace for
  a tree by setting the "default_namespace" attribute and can register
  new prefixes with "register_namespace()".  In XML mode, you can use
  the true/false *xml_declaration* parameter to suppress the XML
  declaration.

* New "Element" method: "extend()" appends the items from a sequence
  to the element’s children.  Elements themselves behave like
  sequences, so it’s easy to move children from one element to
  another:

     from xml.etree import ElementTree as ET

     t = ET.XML("""<list>
       <item>1</item> <item>2</item>  <item>3</item>
     </list>""")
     new = ET.XML('<root/>')
     new.extend(t)

     # Outputs <root><item>1</item>...</root>
     print ET.tostring(new)

* New "Element" method: "iter()" yields the children of the element as
  a generator.  It’s also possible to write "for child in elem:" to
  loop over an element’s children.  The existing method
  "getiterator()" is now deprecated, as is "getchildren()" which
  constructs and returns a list of children.

* New "Element" method: "itertext()" yields all chunks of text that
  are descendants of the element.  For example:

     t = ET.XML("""<list>
       <item>1</item> <item>2</item>  <item>3</item>
     </list>""")

     # Outputs ['\n  ', '1', ' ', '2', '  ', '3', '\n']
     print list(t.itertext())

* Deprecated: using an element as a Boolean (i.e., "if elem:") would
  return true if the element had any children, or false if there were
  no children.  This behaviour is confusing – "None" is false, but so
  is a childless element? – so it will now trigger a "FutureWarning".
  In your code, you should be explicit: write "len(elem) != 0" if
  you’re interested in the number of children, or "elem is not None".

Fredrik Lundh develops ElementTree and produced the 1.3 version; you
can read his article describing 1.3 at https://web.archive.org/web/20
200703234532/http://effbot.org/zone/elementtree-13-intro.htm. Florent
Xicluna updated the version included with Python, after discussions on
python-dev and in bpo-6472.)


Build and C API Changes
=======================

Changes to Python’s build process and to the C API include:

* The latest release of the GNU Debugger, GDB 7, can be scripted using
  Python. When you begin debugging an executable program P, GDB will
  look for a file named "P-gdb.py" and automatically read it.  Dave
  Malcolm contributed a "python-gdb.py" that adds a number of commands
  useful when debugging Python itself.  For example, "py-up" and "py-
  down" go up or down one Python stack frame, which usually
  corresponds to several C stack frames.  "py-print" prints the value
  of a Python variable, and "py-bt" prints the Python stack trace.
  (Added as a result of bpo-8032.)

* If you use the ".gdbinit" file provided with Python, the “pyo” macro
  in the 2.7 version now works correctly when the thread being
  debugged doesn’t hold the GIL; the macro now acquires it before
  printing. (Contributed by Victor Stinner; bpo-3632.)

* "Py_AddPendingCall()" is now thread-safe, letting any worker thread
  submit notifications to the main Python thread.  This is
  particularly useful for asynchronous IO operations. (Contributed by
  Kristján Valur Jónsson; bpo-4293.)

* New function: "PyCode_NewEmpty()" creates an empty code object; only
  the filename, function name, and first line number are required.
  This is useful for extension modules that are attempting to
  construct a more useful traceback stack.  Previously such extensions
  needed to call "PyCode_New()", which had many more arguments.
  (Added by Jeffrey Yasskin.)

* New function: "PyErr_NewExceptionWithDoc()" creates a new exception
  class, just as the existing "PyErr_NewException()" does, but takes
  an extra "char *" argument containing the docstring for the new
  exception class.  (Added by ‘lekma’ on the Python bug tracker;
  bpo-7033.)

* New function: "PyFrame_GetLineNumber()" takes a frame object and
  returns the line number that the frame is currently executing.
  Previously code would need to get the index of the bytecode
  instruction currently executing, and then look up the line number
  corresponding to that address.  (Added by Jeffrey Yasskin.)

* New functions: "PyLong_AsLongAndOverflow()" and
  "PyLong_AsLongLongAndOverflow()"  approximates a Python long integer
  as a C "long" or "long long". If the number is too large to fit into
  the output type, an *overflow* flag is set and returned to the
  caller. (Contributed by Case Van Horsen; bpo-7528 and bpo-7767.)

* New function: stemming from the rewrite of string-to-float
  conversion, a new "PyOS_string_to_double()" function was added.  The
  old "PyOS_ascii_strtod()" and "PyOS_ascii_atof()" functions are now
  deprecated.

* New function: "PySys_SetArgvEx()" sets the value of "sys.argv" and
  can optionally update "sys.path" to include the directory containing
  the script named by "sys.argv[0]" depending on the value of an
  *updatepath* parameter.

  This function was added to close a security hole for applications
  that embed Python.  The old function, "PySys_SetArgv()", would
  always update "sys.path", and sometimes it would add the current
  directory.  This meant that, if you ran an application embedding
  Python in a directory controlled by someone else, attackers could
  put a Trojan-horse module in the directory (say, a file named
  "os.py") that your application would then import and run.

  If you maintain a C/C++ application that embeds Python, check
  whether you’re calling "PySys_SetArgv()" and carefully consider
  whether the application should be using "PySys_SetArgvEx()" with
  *updatepath* set to false.

  Security issue reported as CVE-2008-5983; discussed in bpo-5753, and
  fixed by Antoine Pitrou.

* New macros: the Python header files now define the following macros:
  "Py_ISALNUM", "Py_ISALPHA", "Py_ISDIGIT", "Py_ISLOWER",
  "Py_ISSPACE", "Py_ISUPPER", "Py_ISXDIGIT", "Py_TOLOWER", and
  "Py_TOUPPER". All of these functions are analogous to the C standard
  macros for classifying characters, but ignore the current locale
  setting, because in several places Python needs to analyze
  characters in a locale-independent way.  (Added by Eric Smith;
  bpo-5793.)

* Removed function: "PyEval_CallObject" is now only available as a
  macro.  A function version was being kept around to preserve ABI
  linking compatibility, but that was in 1997; it can certainly be
  deleted by now.  (Removed by Antoine Pitrou; bpo-8276.)

* New format codes: the "PyFormat_FromString()",
  "PyFormat_FromStringV()", and "PyErr_Format()" functions now accept
  "%lld" and "%llu" format codes for displaying C’s "long long" types.
  (Contributed by Mark Dickinson; bpo-7228.)

* The complicated interaction between threads and process forking has
  been changed.  Previously, the child process created by "os.fork()"
  might fail because the child is created with only a single thread
  running, the thread performing the "os.fork()". If other threads
  were holding a lock, such as Python’s import lock, when the fork was
  performed, the lock would still be marked as “held” in the new
  process.  But in the child process nothing would ever release the
  lock, since the other threads weren’t replicated, and the child
  process would no longer be able to perform imports.

  Python 2.7 acquires the import lock before performing an
  "os.fork()", and will also clean up any locks created using the
  "threading" module.  C extension modules that have internal locks,
  or that call "fork()" themselves, will not benefit from this clean-
  up.

  (Fixed by Thomas Wouters; bpo-1590864.)

* The "Py_Finalize()" function now calls the internal
  "threading._shutdown()" function; this prevents some exceptions from
  being raised when an interpreter shuts down. (Patch by Adam Olsen;
  bpo-1722344.)

* When using the "PyMemberDef" structure to define attributes of a
  type, Python will no longer let you try to delete or set a
  "T_STRING_INPLACE" attribute.

* Global symbols defined by the "ctypes" module are now prefixed with
  "Py", or with "_ctypes".  (Implemented by Thomas Heller; bpo-3102.)

* New configure option: the "--with-system-expat" switch allows
  building the "pyexpat" module to use the system Expat library.
  (Contributed by Arfrever Frehtes Taifersar Arahesis; bpo-7609.)

* New configure option: the "--with-valgrind" option will now disable
  the pymalloc allocator, which is difficult for the Valgrind memory-
  error detector to analyze correctly. Valgrind will therefore be
  better at detecting memory leaks and overruns. (Contributed by James
  Henstridge; bpo-2422.)

* New configure option: you can now supply an empty string to "--with-
  dbmliborder=" in order to disable all of the various DBM modules.
  (Added by Arfrever Frehtes Taifersar Arahesis; bpo-6491.)

* The **configure** script now checks for floating-point rounding bugs
  on certain 32-bit Intel chips and defines a "X87_DOUBLE_ROUNDING"
  preprocessor definition.  No code currently uses this definition,
  but it’s available if anyone wishes to use it. (Added by Mark
  Dickinson; bpo-2937.)

  **configure** also now sets a "LDCXXSHARED" Makefile variable for
  supporting C++ linking.  (Contributed by Arfrever Frehtes Taifersar
  Arahesis; bpo-1222585.)

* The build process now creates the necessary files for pkg-config
  support.  (Contributed by Clinton Roy; bpo-3585.)

* The build process now supports Subversion 1.7.  (Contributed by
  Arfrever Frehtes Taifersar Arahesis; bpo-6094.)


Capsules
--------

Python 3.1 adds a new C datatype, "PyCapsule", for providing a C API
to an extension module.  A capsule is essentially the holder of a C
"void *" pointer, and is made available as a module attribute; for
example, the "socket" module’s API is exposed as "socket.CAPI", and
"unicodedata" exposes "ucnhash_CAPI".  Other extensions can import the
module, access its dictionary to get the capsule object, and then get
the "void *" pointer, which will usually point to an array of pointers
to the module’s various API functions.

There is an existing data type already used for this, "PyCObject", but
it doesn’t provide type safety.  Evil code written in pure Python
could cause a segmentation fault by taking a "PyCObject" from module A
and somehow substituting it for the "PyCObject" in module B.
Capsules know their own name, and getting the pointer requires
providing the name:

   void *vtable;

   if (!PyCapsule_IsValid(capsule, "mymodule.CAPI") {
           PyErr_SetString(PyExc_ValueError, "argument type invalid");
           return NULL;
   }

   vtable = PyCapsule_GetPointer(capsule, "mymodule.CAPI");

You are assured that "vtable" points to whatever you’re expecting. If
a different capsule was passed in, "PyCapsule_IsValid()" would detect
the mismatched name and return false.  Refer to Providing a C API for
an Extension Module for more information on using these objects.

Python 2.7 now uses capsules internally to provide various extension-
module APIs, but the "PyCObject_AsVoidPtr()" was modified to handle
capsules, preserving compile-time compatibility with the "CObject"
interface.  Use of "PyCObject_AsVoidPtr()" will signal a
"PendingDeprecationWarning", which is silent by default.

Implemented in Python 3.1 and backported to 2.7 by Larry Hastings;
discussed in bpo-5630.


Port-Specific Changes: Windows
------------------------------

* The "msvcrt" module now contains some constants from the
  "crtassem.h" header file: "CRT_ASSEMBLY_VERSION",
  "VC_ASSEMBLY_PUBLICKEYTOKEN", and "LIBRARIES_ASSEMBLY_NAME_PREFIX".
  (Contributed by David Cournapeau; bpo-4365.)

* The "_winreg" module for accessing the registry now implements the
  "CreateKeyEx()" and "DeleteKeyEx()" functions, extended versions of
  previously-supported functions that take several extra arguments.
  The "DisableReflectionKey()", "EnableReflectionKey()", and
  "QueryReflectionKey()" were also tested and documented. (Implemented
  by Brian Curtin: bpo-7347.)

* The new "_beginthreadex()" API is used to start threads, and the
  native thread-local storage functions are now used. (Contributed by
  Kristján Valur Jónsson; bpo-3582.)

* The "os.kill()" function now works on Windows.  The signal value can
  be the constants "CTRL_C_EVENT", "CTRL_BREAK_EVENT", or any integer.
  The first two constants will send "Control-C" and "Control-Break"
  keystroke events to subprocesses; any other value will use the
  "TerminateProcess()" API.  (Contributed by Miki Tebeka;
  bpo-1220212.)

* The "os.listdir()" function now correctly fails for an empty path.
  (Fixed by Hirokazu Yamamoto; bpo-5913.)

* The "mimelib" module will now read the MIME database from the
  Windows registry when initializing. (Patch by Gabriel Genellina;
  bpo-4969.)


Port-Specific Changes: Mac OS X
-------------------------------

* The path "/Library/Python/2.7/site-packages" is now appended to
  "sys.path", in order to share added packages between the system
  installation and a user-installed copy of the same version. (Changed
  by Ronald Oussoren; bpo-4865.)

     Changed in version 2.7.13: As of 2.7.13, this change was removed.
     "/Library/Python/2.7/site-packages", the site-packages directory
     used by the Apple-supplied system Python 2.7 is no longer
     appended to "sys.path" for user-installed Pythons such as from
     the python.org installers.  As of macOS 10.12, Apple changed how
     the system site-packages directory is configured, which could
     cause installation of pip components, like setuptools, to fail.
     Packages installed for the system Python will no longer be shared
     with user-installed Pythons. (bpo-28440)


Port-Specific Changes: FreeBSD
------------------------------

* FreeBSD 7.1’s "SO_SETFIB" constant, used with
  "getsockopt()"/"setsockopt()" to select an alternate routing table,
  is now available in the "socket" module.  (Added by Kyle VanderBeek;
  bpo-8235.)


Other Changes and Fixes
=======================

* Two benchmark scripts, "iobench" and "ccbench", were added to the
  "Tools" directory.  "iobench" measures the speed of the built-in
  file I/O objects returned by "open()" while performing various
  operations, and "ccbench" is a concurrency benchmark that tries to
  measure computing throughput, thread switching latency, and IO
  processing bandwidth when performing several tasks using a varying
  number of threads.

* The "Tools/i18n/msgfmt.py" script now understands plural forms in
  ".po" files.  (Fixed by Martin von Löwis; bpo-5464.)

* When importing a module from a ".pyc" or ".pyo" file with an
  existing ".py" counterpart, the "co_filename" attributes of the
  resulting code objects are overwritten when the original filename is
  obsolete.  This can happen if the file has been renamed, moved, or
  is accessed through different paths.  (Patch by Ziga Seilnacht and
  Jean-Paul Calderone; bpo-1180193.)

* The "regrtest.py" script now takes a "--randseed=" switch that takes
  an integer that will be used as the random seed for the "-r" option
  that executes tests in random order. The "-r" option also reports
  the seed that was used (Added by Collin Winter.)

* Another "regrtest.py" switch is "-j", which takes an integer
  specifying how many tests run in parallel. This allows reducing the
  total runtime on multi-core machines. This option is compatible with
  several other options, including the "-R" switch which is known to
  produce long runtimes. (Added by Antoine Pitrou, bpo-6152.)  This
  can also be used with a new "-F" switch that runs selected tests in
  a loop until they fail.  (Added by Antoine Pitrou; bpo-7312.)

* When executed as a script, the "py_compile.py" module now accepts
  "'-'" as an argument, which will read standard input for the list of
  filenames to be compiled.  (Contributed by Piotr Ożarowski;
  bpo-8233.)


Porting to Python 2.7
=====================

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

* The "range()" function processes its arguments more consistently; it
  will now call "__int__()" on non-float, non-integer arguments that
  are supplied to it.  (Fixed by Alexander Belopolsky; bpo-1533.)

* The string "format()" method changed the default precision used for
  floating-point and complex numbers from 6 decimal places to 12,
  which matches the precision used by "str()". (Changed by Eric Smith;
  bpo-5920.)

* Because of an optimization for the "with" statement, the special
  methods "__enter__()" and "__exit__()" must belong to the object’s
  type, and cannot be directly attached to the object’s instance.
  This affects new-style classes (derived from "object") and C
  extension types.  (bpo-6101.)

* Due to a bug in Python 2.6, the *exc_value* parameter to
  "__exit__()" methods was often the string representation of the
  exception, not an instance.  This was fixed in 2.7, so *exc_value*
  will be an instance as expected.  (Fixed by Florent Xicluna;
  bpo-7853.)

* When a restricted set of attributes were set using "__slots__",
  deleting an unset attribute would not raise "AttributeError" as you
  would expect.  Fixed by Benjamin Peterson; bpo-7604.)

In the standard library:

* Operations with "datetime" instances that resulted in a year falling
  outside the supported range didn’t always raise "OverflowError".
  Such errors are now checked more carefully and will now raise the
  exception. (Reported by Mark Leander, patch by Anand B. Pillai and
  Alexander Belopolsky; bpo-7150.)

* When using "Decimal" instances with a string’s "format()" method,
  the default alignment was previously left-alignment.  This has been
  changed to right-alignment, which might change the output of your
  programs. (Changed by Mark Dickinson; bpo-6857.)

  Comparisons involving a signaling NaN value (or "sNAN") now signal
  "InvalidOperation" instead of silently returning a true or false
  value depending on the comparison operator.  Quiet NaN values (or
  "NaN") are now hashable.  (Fixed by Mark Dickinson; bpo-7279.)

* The ElementTree library, "xml.etree", no longer escapes ampersands
  and angle brackets when outputting an XML processing instruction
  (which looks like *<?xml-stylesheet href=”#style1”?>*) or comment
  (which looks like *<!– comment –>*). (Patch by Neil Muller;
  bpo-2746.)

* The "readline()" method of "StringIO" objects now does nothing when
  a negative length is requested, as other file-like objects do.
  (bpo-7348).

* The "syslog" module will now use the value of "sys.argv[0]" as the
  identifier instead of the previous default value of "'python'".
  (Changed by Sean Reifschneider; bpo-8451.)

* The "tarfile" module’s default error handling has changed, to no
  longer suppress fatal errors.  The default error level was
  previously 0, which meant that errors would only result in a message
  being written to the debug log, but because the debug log is not
  activated by default, these errors go unnoticed.  The default error
  level is now 1, which raises an exception if there’s an error.
  (Changed by Lars Gustäbel; bpo-7357.)

* The "urlparse" module’s "urlsplit()" now handles unknown URL schemes
  in a fashion compliant with **RFC 3986**: if the URL is of the form
  ""<something>://..."", the text before the "://" is treated as the
  scheme, even if it’s a made-up scheme that the module doesn’t know
  about.  This change may break code that worked around the old
  behaviour.  For example, Python 2.6.4 or 2.5 will return the
  following:

     >>> import urlparse
     >>> urlparse.urlsplit('invented://host/filename?query')
     ('invented', '', '//host/filename?query', '', '')

  Python 2.7 (and Python 2.6.5) will return:

     >>> import urlparse
     >>> urlparse.urlsplit('invented://host/filename?query')
     ('invented', 'host', '/filename?query', '', '')

  (Python 2.7 actually produces slightly different output, since it
  returns a named tuple instead of a standard tuple.)

For C extensions:

* C extensions that use integer format codes with the "PyArg_Parse*"
  family of functions will now raise a "TypeError" exception instead
  of triggering a "DeprecationWarning" (bpo-5080).

* Use the new "PyOS_string_to_double()" function instead of the old
  "PyOS_ascii_strtod()" and "PyOS_ascii_atof()" functions, which are
  now deprecated.

For applications that embed Python:

* The "PySys_SetArgvEx()" function was added, letting applications
  close a security hole when the existing "PySys_SetArgv()" function
  was used.  Check whether you’re calling "PySys_SetArgv()" and
  carefully consider whether the application should be using
  "PySys_SetArgvEx()" with *updatepath* set to false.


New Features Added to Python 2.7 Maintenance Releases
=====================================================

New features may be added to Python 2.7 maintenance releases when the
situation genuinely calls for it. Any such additions must go through
the Python Enhancement Proposal process, and make a compelling case
for why they can’t be adequately addressed by either adding the new
feature solely to Python 3, or else by publishing it on the Python
Package Index.

In addition to the specific proposals listed below, there is a general
exemption allowing new "-3" warnings to be added in any Python 2.7
maintenance release.


Two new environment variables for debug mode
--------------------------------------------

In debug mode, the "[xxx refs]" statistic is not written by default,
the "PYTHONSHOWREFCOUNT" environment variable now must also be set.
(Contributed by Victor Stinner; bpo-31733.)

When Python is compiled with "COUNT_ALLOC" defined, allocation counts
are no longer dumped by default anymore: the "PYTHONSHOWALLOCCOUNT"
environment variable must now also be set. Moreover, allocation counts
are now dumped into stderr, rather than stdout. (Contributed by Victor
Stinner; bpo-31692.)

New in version 2.7.15.


PEP 434: IDLE Enhancement Exception for All Branches
----------------------------------------------------

**PEP 434** describes a general exemption for changes made to the IDLE
development environment shipped along with Python. This exemption
makes it possible for the IDLE developers to provide a more consistent
user experience across all supported versions of Python 2 and 3.

For details of any IDLE changes, refer to the NEWS file for the
specific release.


PEP 466: Network Security Enhancements for Python 2.7
-----------------------------------------------------

**PEP 466** describes a number of network security enhancement
proposals that have been approved for inclusion in Python 2.7
maintenance releases, with the first of those changes appearing in the
Python 2.7.7 release.

**PEP 466** related features added in Python 2.7.7:

* "hmac.compare_digest()" was backported from Python 3 to make a
  timing attack resistant comparison operation available to Python 2
  applications. (Contributed by Alex Gaynor; bpo-21306.)

* OpenSSL 1.0.1g was upgraded in the official Windows installers
  published on python.org. (Contributed by Zachary Ware; bpo-21462.)

**PEP 466** related features added in Python 2.7.8:

* "hashlib.pbkdf2_hmac()" was backported from Python 3 to make a
  hashing algorithm suitable for secure password storage broadly
  available to Python 2 applications. (Contributed by Alex Gaynor;
  bpo-21304.)

* OpenSSL 1.0.1h was upgraded for the official Windows installers
  published on python.org. (contributed by Zachary Ware in bpo-21671
  for CVE-2014-0224)

**PEP 466** related features added in Python 2.7.9:

* Most of Python 3.4’s "ssl" module was backported. This means "ssl"
  now supports Server Name Indication, TLS1.x settings, access to the
  platform certificate store, the "SSLContext" class, and other
  features. (Contributed by Alex Gaynor and David Reid; bpo-21308.)

  Refer to the “Version added: 2.7.9” notes in the module
  documentation for specific details.

* "os.urandom()" was changed to cache a file descriptor to
  "/dev/urandom" instead of reopening "/dev/urandom" on every call.
  (Contributed by Alex Gaynor; bpo-21305.)

* "hashlib.algorithms_guaranteed" and "hashlib.algorithms_available"
  were backported from Python 3 to make it easier for Python 2
  applications to select the strongest available hash algorithm.
  (Contributed by Alex Gaynor in bpo-21307)


PEP 477: Backport ensurepip (PEP 453) to Python 2.7
---------------------------------------------------

**PEP 477** approves the inclusion of the **PEP 453** ensurepip module
and the improved documentation that was enabled by it in the Python
2.7 maintenance releases, appearing first in the Python 2.7.9 release.


Bootstrapping pip By Default
~~~~~~~~~~~~~~~~~~~~~~~~~~~~

The new "ensurepip" module (defined in **PEP 453**) provides a
standard cross-platform mechanism to bootstrap the pip installer into
Python installations. The version of "pip" included with Python 2.7.9
is "pip" 1.5.6, and future 2.7.x maintenance releases will update the
bundled version to the latest version of "pip" that is available at
the time of creating the release candidate.

By default, the commands "pip", "pipX" and "pipX.Y" will be installed
on all platforms (where X.Y stands for the version of the Python
installation), along with the "pip" Python package and its
dependencies.

For CPython source builds on POSIX systems, the "make install" and
"make altinstall" commands do not bootstrap "pip" by default.  This
behaviour can be controlled through configure options, and overridden
through Makefile options.

On Windows and Mac OS X, the CPython installers now default to
installing "pip" along with CPython itself (users may opt out of
installing it during the installation process). Window users will need
to opt in to the automatic "PATH" modifications to have "pip"
available from the command line by default, otherwise it can still be
accessed through the Python launcher for Windows as "py -m pip".

As discussed in the PEP, platform packagers may choose not to install
these commands by default, as long as, when invoked, they provide
clear and simple directions on how to install them on that platform
(usually using the system package manager).


Documentation Changes
~~~~~~~~~~~~~~~~~~~~~

As part of this change, the Installing Python Modules and Distributing
Python Modules sections of the documentation have been completely
redesigned as short getting started and FAQ documents. Most packaging
documentation has now been moved out to the Python Packaging Authority
maintained Python Packaging User Guide and the documentation of the
individual projects.

However, as this migration is currently still incomplete, the legacy
versions of those guides remaining available as Installing Python
Modules (Legacy version) and Distributing Python Modules (Legacy
version).

See also:

  **PEP 453** – Explicit bootstrapping of pip in Python installations
     PEP written by Donald Stufft and Nick Coghlan, implemented by
     Donald Stufft, Nick Coghlan, Martin von Löwis and Ned Deily.


PEP 476: Enabling certificate verification by default for stdlib http clients
-----------------------------------------------------------------------------

**PEP 476** updated "httplib" and modules which use it, such as
"urllib2" and "xmlrpclib", to now verify that the server presents a
certificate which is signed by a Certificate Authority in the platform
trust store and whose hostname matches the hostname being requested by
default, significantly improving security for many applications. This
change was made in the Python 2.7.9 release.

For applications which require the old previous behavior, they can
pass an alternate context:

   import urllib2
   import ssl

   # This disables all verification
   context = ssl._create_unverified_context()

   # This allows using a specific certificate for the host, which doesn't need
   # to be in the trust store
   context = ssl.create_default_context(cafile="/path/to/file.crt")

   urllib2.urlopen("https://invalid-cert", context=context)


PEP 493: HTTPS verification migration tools for Python 2.7
----------------------------------------------------------

**PEP 493** provides additional migration tools to support a more
incremental infrastructure upgrade process for environments containing
applications and services relying on the historically permissive
processing of server certificates when establishing client HTTPS
connections.  These additions were made in the Python 2.7.12 release.

These tools are intended for use in cases where affected applications
and services can’t be modified to explicitly pass a more permissive
SSL context when establishing the connection.

For applications and services which can’t be modified at all, the new
"PYTHONHTTPSVERIFY" environment variable may be set to "0" to revert
an entire Python process back to the default permissive behaviour of
Python 2.7.8 and earlier.

For cases where the connection establishment code can’t be modified,
but the overall application can be, the new
"ssl._https_verify_certificates()" function can be used to adjust the
default behaviour at runtime.


New "make regen-all" build target
---------------------------------

To simplify cross-compilation, and to ensure that CPython can reliably
be compiled without requiring an existing version of Python to already
be available, the autotools-based build system no longer attempts to
implicitly recompile generated files based on file modification times.

Instead, a new "make regen-all" command has been added to force
regeneration of these files when desired (e.g. after an initial
version of Python has already been built based on the pregenerated
versions).

More selective regeneration targets are also defined - see
Makefile.pre.in for details.

(Contributed by Victor Stinner in bpo-23404.)

New in version 2.7.14.


Removal of "make touch" build target
------------------------------------

The "make touch" build target previously used to request implicit
regeneration of generated files by updating their modification times
has been removed.

It has been replaced by the new "make regen-all" target.

(Contributed by Victor Stinner in bpo-23404.)

Changed in version 2.7.14.


Acknowledgements
================

The author would like to thank the following people for offering
suggestions, corrections and assistance with various drafts of this
article: Nick Coghlan, Philip Jenvey, Ryan Lovett, R. David Murray,
Hugh Secker-Walker.
