What's New In Python 3.4
This article explains the new features in Python 3.4, compared to 3.3.
For full details, see the changelog.
Prerelease users should be aware that this document is currently in draft form. It will be updated substantially as Python 3.4 moves towards release, so it's worth checking back even after reading earlier versions.
Summary -- Release highlights
New syntax features:
- No new syntax features were added in Python 3.4.
New library modules:
- :mod:`asyncio`: New provisonal API for asynchronous IO (PEP 3156).
- :mod:`enum`: Support for enumeration types (PEP 435).
- :mod:`ensurepip`: Bootstrapping the pip installer (PEP 453).
- :mod:`pathlib`: Object-oriented filesystem paths (PEP 428).
- :mod:`selectors`: High-level and efficient I/O multiplexing, built upon the :mod:`select` module primitives.
- :mod:`statistics`: A basic numerically stable statistics library (PEP 450).
- :mod:`tracemalloc`: Trace Python memory allocations (PEP 454).
New expected features for Python implementations:
- PEP 446: :ref:`Make newly created file descriptors non-inheritable <whatsnew-pep-446>`.
- command line option for :ref:`isolated mode <using-on-misc-options>`, (:issue:`16499`).
- :ref:`improvements <codec-handling-improvements>` in the handling of codecs that are not text encodings
Significantly Improved Library Modules:
- :ref:`Single-dispatch generic functions <whatsnew-singledispatch>` in :mod:`functools` (PEP 443)
- New :mod:`pickle` :ref:`protocol 4 <whatsnew-protocol-4>` (PEP 3154)
- :ref:`SHA-3 (Keccak) support <whatsnew-sha3>` for :mod:`hashlib`.
- :ref:`TLSv1.1 and TLSv1.2 support <whatsnew-tls-11-12>` for :mod:`ssl`.
- :mod:`multiprocessing` now has :ref:`an option to avoid using os.fork on Unix <whatsnew-multiprocessing-no-fork>` (:issue:`8713`).
- :mod:`email` has a new submodule, :mod:`~email.contentmanager`, and a new :mod:`~email.message.Message` subclass (:class:`~email.contentmanager.EmailMessage`) that :ref:`simplify MIME handling <whatsnew_email_contentmanager>`.
CPython implementation improvements:
- PEP 442: :ref:`Safe object finalization <whatsnew-pep-442>`
- Leveraging PEP 442, :ref:`module globals are no longer set to None during finalization <whatsnew-pep-442>`, in most cases (:issue:`18214`).
- PEP 445: :ref:`Configurable memory allocators <whatsnew-pep-445>`
- PEP 456: :ref:`Secure and interchangeable hash algorithm <whatsnew-pep-456>`
- PEP 436: :ref:`Argument Clinic <whatsnew-pep-436>`.
- A more efficient :mod:`marshal` format (:issue:`16475`).
Please read on for a comprehensive list of user-facing changes.
PEP 453: Explicit bootstrapping of pip in Python installations
The :mod:`venv` module and the :command:`pyvenv` utility make use of this module to make pip readily available in virtual environments. When using the command line interface, pip is installed by default, while for the module API installation of pip must be requested explicitly.
For CPython source builds on POSIX systems, the make install and make altinstall commands 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 offer the option to install pip along with CPython itself.
The implementation of PEP 453 is still a work in progress. Refer to :issue:`19347` for the progress on additional steps:
- Having the binary installers install pip by default
- Recommending the use of pip in the "Installing Python Module" documentation.
PEP 446: Make newly created file descriptors non-inheritable
Improvements to codec handling
Since it was first introduced, the :mod:`codecs` module has always been intended to operate as a type-neutral dynamic encoding and decoding system. However, its close coupling with the Python text model, especially the type restricted convenience methods on the builtin :class:`str`, :class:`bytes` and :class:`bytearray` types, has historically obscured that fact.
As a key step in clarifying the situation, the :meth:`codecs.encode` and :meth:`codecs.decode` convenience functions are now properly documented in Python 2.7, 3.3 and 3.4. These functions have existed in the :mod:`codecs` module (and have been covered by the regression test suite) since Python 2.4, but were previously only discoverable through runtime introspection.
Unlike the convenience methods on :class:`str`, :class:`bytes` and :class:`bytearray`, these convenience functions support arbitrary codecs in both Python 2 and Python 3, rather than being limited to Unicode text encodings (in Python 3) or basestring <-> basestring conversions (in Python 2).
In Python 3.4, the interpreter is able to identify the known non-text encodings provided in the standard library and direct users towards these general purpose convenience functions when appropriate:
>>> b"abcdef".decode("hex") Traceback (most recent call last): File "<stdin>", line 1, in <module> LookupError: 'hex' is not a text encoding; use codecs.decode() to handle arbitrary codecs >>> "hello".encode("rot13") Traceback (most recent call last): File "<stdin>", line 1, in <module> LookupError: 'rot13' is not a text encoding; use codecs.encode() to handle arbitrary codecs
In a related change, whenever it is feasible without breaking backwards compatibility, exceptions raised during encoding and decoding operations will be wrapped in a chained exception of the same type that mentions the name of the codec responsible for producing the error:
>>> import codecs >>> codecs.decode(b"abcdefgh", "hex") binascii.Error: Non-hexadecimal digit found The above exception was the direct cause of the following exception: Traceback (most recent call last): File "<stdin>", line 1, in <module> binascii.Error: decoding with 'hex' codec failed (Error: Non-hexadecimal digit found) >>> codecs.encode("hello", "bz2") TypeError: 'str' does not support the buffer interface The above exception was the direct cause of the following exception: Traceback (most recent call last): File "<stdin>", line 1, in <module> TypeError: encoding with 'bz2' codec failed (TypeError: 'str' does not support the buffer interface)
Finally, as the examples above show, these improvements have permitted the restoration of the convenience aliases for the non-Unicode codecs that were themselves restored in Python 3.2. This means that encoding binary data to and from its hexadecimal representation (for example) can now be written as:
>>> from codecs import encode, decode >>> encode(b"hello", "hex") b'68656c6c6f' >>> decode(b"68656c6c6f", "hex") b'hello'
PEP 451: A ModuleSpec Type for the Import System
PEP 451 provides an encapsulation of the information about a module that the import machinery will use to load it, (i.e. a module spec). This helps simplify both the import implementation and several import-related APIs. The change is also a stepping stone for several future import-related improvements.
The public-facing changes from the PEP are entirely backward-compatible. Furthermore, they should be transparent to everyone but importer authors. Key finder and loader methods have been deprecated, but they will continue working. New importers should use the new methods described in the PEP. Existing importers should be updated to implement the new methods.
Pickle protocol 4
The new :mod:`pickle` protocol addresses a number of issues that were present in previous protocols, such as the serialization of nested classes, very large strings and containers, or classes whose :meth:`__new__` method takes keyword-only arguments. It also brings a couple efficiency improvements.
Other Language Changes
Some smaller changes made to the core Python language are:
- Unicode database updated to UCD version 6.3.
- :func:`min` and :func:`max` now accept a default argument that can be used to specify the value they return if the iterable they are evaluating has no elements. Contributed by Julian Berman in :issue:`18111`.
- Module objects are now :mod:`weakref`'able.
- Module __file__ attributes (and related values) should now always contain absolute paths by default, with the sole exception of __main__.__file__ when a script has been executed directly using a relative path (Contributed by Brett Cannon in :issue:`18416`).
- Now all the UTF-* codecs (except UTF-7) reject surrogates during both encoding and decoding unless the surrogatepass error handler is used, with the exception of the UTF-16 decoder that accepts valid surrogate pairs, and the UTF-16 encoder that produces them while encoding non-BMP characters. Contributed by Victor Stinner, Kang-Hao (Kenny) Lu and Serhiy Storchaka in :issue:`12892`.
The new :mod:`asyncio` module (defined in PEP 3156) provides a standard pluggable event loop model for Python, providing solid asynchronous IO support in the standard library, and making it easier for other event loop implementations to interoperate with the standard library and each other.
For Python 3.4, this module is considered a :term:`provisional API`.
The new :mod:`enum` module (defined in PEP 435) provides a standard implementation of enumeration types, allowing other modules (such as :mod:`socket`) to provide more informative error messages and better debugging support by replacing opaque integer constants with backwards compatible enumeration values.
The new :mod:`pathlib` module offers classes representing filesystem paths with semantics appropriate for different operating systems. Path classes are divided between pure paths, which provide purely computational operations without I/O, and concrete paths, which inherit from pure paths but also provide I/O operations.
For Python 3.4, this module is considered a :term:`provisional API`.
The new :mod:`statistics` module (defined in PEP 450) offers some core statistics functionality directly in the standard library. This module supports calculation of the mean, median, mode, variance and standard deviation of a data series.
- Traceback where an object was allocated
- Statistics on allocated memory blocks per filename and per line number: total size, number and average size of allocated memory blocks
- Compute the differences between two snapshots to detect memory leaks
Added support for 24-bit samples (:issue:`12866`).
The number of digits in the coefficients for the RGB --- YIQ conversions have been expanded so that they match the FCC NTSC versions. The change in results should be less than 1% and may better match results found elsewhere.
The new :class:`contextlib.suppress` context manager helps to clarify the intent of code that deliberately suppresses exceptions from a single statement. (Contributed by Raymond Hettinger in :issue:`15806` and Zero Piraeus in :issue:`19266`)
The new :func:`contextlib.redirect_stdout` context manager makes it easier for utility scripts to handle inflexible APIs that don't provide any options to retrieve their output as a string or direct it to somewhere other than :data:`sys.stdout`. In conjunction with :class:`io.StringIO`, this context manager is also useful for checking expected output from command line utilities. (Contribute by Raymond Hettinger in :issue:`15805`)
The :mod:`contextlib` documentation has also been updated to include a :ref:`discussion <single-use-reusable-and-reentrant-cms>` of the differences between single use, reusable and reentrant context managers.
The :mod:`dis` module is now built around an :class:`~dis.Instruction` class that provides details of individual bytecode operations and a :func:`~dis.get_instructions` iterator that emits the Instruction stream for a given piece of Python code. The various display tools in the :mod:`dis` module have been updated to be based on these new components.
The new :class:`dis.Bytecode` class provides an object-oriented API for inspecting bytecode, both in human-readable form and for iterating over instructions.
Updated the doctest command line interface to use :mod:`argparse`, and added -o and -f options to the interface. -o allows doctest options to be specified on the command line, and -f is a shorthand for -o FAIL_FAST (to parallel the similar option supported by the :mod:`unittest` CLI). (Contributed by R. David Murray in :issue:`11390`.)
:meth:`~email.message.Message.as_string` now accepts a policy argument to override the default policy of the message when generating a string representation of it. This means that as_string can now be used in more circumstances, instead of having to create and use a :mod:`~email.generator` in order to pass formatting parameters to its flatten method.
New method :meth:`~email.message.Message.as_bytes` added to produce a bytes representation of the message in a fashion similar to how as_string produces a string representation. It does not accept the maxheaderlen argument, but does accept the unixfrom and policy arguments. The :class:`~email.message.Message` :meth:`~email.message.Message.__bytes__` method calls it, meaning that bytes(mymsg) will now produce the intuitive result: a bytes object containing the fully formatted message.
(Contributed by R. David Murray in :issue:`18600`.)
A pair of new subclasses of :class:`~email.message.Message` have been added, along with a new sub-module, :mod:`~email.contentmanager`. All documentation is currently in the new module, which is being added as part of the new :term:`provisional <provisional package>` email API. These classes provide a number of new methods that make extracting content from and inserting content into email messages much easier. See the :mod:`~email.contentmanager` documentation for details.
These API additions complete the bulk of the work that was planned as part of the email6 project. The currently provisional API is scheduled to become final in Python 3.5 (possibly with a few minor additions in the area of error handling).
(Contributed by R. David Murray in :issue:`18891`.)
The new :func:`~functools.partialmethod` descriptor bring partial argument application to descriptors, just as :func:`~functools.partial` provides for normal callables. The new descriptor also makes it easier to get arbitrary callables (including :func:`~functools.partial` instances) to behave like normal instance methods when included in a class definition.
(Contributed by Alon Horev and Nick Coghlan in :issue:`4331`)
The new :func:`~functools.singledispatch` decorator brings support for single-dispatch generic functions to the Python standard library. Where object oriented programming focuses on grouping multiple operations on a common set of data into a class, a generic function focuses on grouping multiple implementations of an operation that allows it to work with different kinds of data.
Added a new convert_charrefs keyword argument to :class:`~html.parser.HTMLParser` that, when True, automatically converts all character references. For backward-compatibility, its value defaults to False, but it will change to True in future versions, so you are invited to set it explicitly and update your code to use this new feature. (Contributed by Ezio Melotti in :issue:`13633`)
The inspect module now offers a basic :ref:`command line interface <inspect-module-cli>` to quickly display source code and other information for modules, classes and functions. (Contributed by Claudiu Popa and Nick Coghlan in :issue:`18626`)
:func:`~inspect.unwrap` makes it easy to unravel wrapper function chains created by :func:`functools.wraps` (and any other API that sets the __wrapped__ attribute on a wrapper function). (Contributed by Daniel Urban, Aaron Iles and Nick Coghlan in :issue:`13266`)
As part of the implementation of the new :mod:`enum` module, the :mod:`inspect` module now has substantially better support for custom __dir__ methods and dynamic class attributes provided through metaclasses (Contributed by Ethan Furman in :issue:`18929` and :issue:`19030`)
mmap objects can now be weakref'ed. (Contributed by Valerie Lambert in :issue:`4885`.)
On Unix, two new :ref:`start methods <multiprocessing-start-methods>` (spawn and forkserver) have been added for starting processes using :mod:`multiprocessing`. These make the mixing of processes with threads more robust, and the spawn method matches the semantics that multiprocessing has always used on Windows. (Contributed by Richard Oudkerk in :issue:`8713`).
Also, except when using the old fork start method, child processes will no longer inherit unneeded handles/file descriptors from their parents (part of :issue:`8713`).
:mod:`multiprocessing` now relies on :mod:`runpy` (which implements the -m switch) to initialise __main__ appropriately in child processes when using the spawn or forkserver start methods. This resolves some edge cases where combining multiprocessing, the -m command line switch and explicit relative imports could cause obscure failures in child processes. (Contributed by Nick Coghlan in :issue:`19946`)
New functions to get and set the :ref:`inheritable flag <fd_inheritance>` of a file descriptors or a Windows handle:
The print command has been removed from :mod:`pdb`, restoring access to the print function.
Rationale: Python2's pdb did not have a print command; instead, entering print executed the print statement. In Python3 print was mistakenly made an alias for the pdb :pdbcmd:`p` command. p, however, prints the repr of its argument, not the str like the Python2 print command did. Worse, the Python3 pdb print command shadowed the Python3 print function, making it inaccessible at the pdb prompt.
(Contributed by Connor Osborn in :issue:`18764`.)
New :meth:`~poplib.POP3.stls` method to switch a clear-text POP3 session into an encrypted POP3 session.
New :meth:`~poplib.POP3.capa` method to query the capabilities advertised by the POP3 server.
(Contributed by Lorenzo Catucci in :issue:`4473`.)
While significant changes have not been made to :mod:`pydoc` directly, its handling of custom __dir__ methods and various descriptor behaviours has been improved substantially by the underlying changes in the :mod:`inspect` module.
The repr of :ref:`regex objects <re-objects>` now includes the pattern and the flags; the repr of :ref:`match objects <match-objects>` now includes the start, end, and the part of the string that matched. (Contributed by Serhiy Storchaka in :issue:`13592` and :issue:`17087`.)
:exc:`~smtplib.SMTPException` is now a subclass of :exc:`OSError`, which allows both socket level errors and SMTP protocol level errors to be caught in one try/except statement by code that only cares whether or not an error occurred. (:issue:`2118`).
Socket objects have new methods to get or set their :ref:`inheritable flag <fd_inheritance>`:
The socket.AF_* and socket.SOCK_* constants are enumeration values, using the new :mod:`enum` module. This allows descriptive reporting during debugging, instead of seeing integer "magic numbers".
:data:`~ssl.PROTOCOL_TLSv1_1` and :data:`~ssl.PROTOCOL_TLSv1_2` (TLSv1.1 and TLSv1.2 support) have been added; support for these protocols is only available if Python is linked with OpenSSL 1.0.1 or later. (Contributed by Michele Orrù and Antoine Pitrou in :issue:`16692`)
New diagnostic functions :func:`~ssl.get_default_verify_paths`, :meth:`~ssl.SSLContext.cert_store_stats` and :meth:`~ssl.SSLContext.get_ca_certs` (Contributed by Christian Heimes in :issue:`18143` and :issue:`18147`)
The :mod:`stat` module is now backed by a C implementation in :mod:`_stat`. A C implementation is required as most of the values aren't standardized and platform-dependent. (Contributed by Christian Heimes in :issue:`11016`.)
The module supports new file types: door, event port and whiteout.
New :class:`~weakref.finalize` class makes it possible to register a callback to be invoked when an object is garbage collected, without needing to carefully manage the lifecycle of the weak reference itself. (Contributed by Richard Oudkerk in :issue:`15528`)
Tab-completion is now enabled by default in the interactive interpreter. (Contributed by Antoine Pitrou and Éric Araujo in :issue:`5845`.)
Python invocation changes
Invoking the Python interpreter with --version now outputs the version to standard output instead of standard error (:issue:`18338`). Similar changes were made to :mod:`argparse` (:issue:`18920`) and other modules that have script-like invocation capabilities (:issue:`18922`).
Major performance enhancements have been added:
The UTF-32 decoder is now 3x to 4x faster.
The cost of hash collisions for sets is now reduced. Each hash table probe now checks a series of consecutive, adjacent key/hash pairs before continuing to make random probes through the hash table. This exploits cache locality to make collision resolution less expensive.
The collision resolution scheme can be described as a hybrid of linear probing and open addressing. The number of additional linear probes defaults to nine. This can be changed at compile-time by defining LINEAR_PROBES to be any value. Set LINEAR_PROBES=0 to turn-off linear probing entirely.
(Contributed by Raymond Hettinger in :issue:`18771`.)
The interpreter starts about 30% faster. A couple of measures lead to the speedup. The interpreter loads fewer modules on startup, e.g. the :mod:`re`, :mod:`collections` and :mod:`locale` modules and their dependencies are no longer imported by default. The marshal module has been improved to load compiled Python code faster.
CPython Implementation Changes
PEP 445: Customization of CPython memory allocators
PEP 445 adds new C level interfaces to customize memory allocation in the CPython interpreter.
PEP 442: Safe object finalization
PEP 442 removes the current limitations and quirks of object finalization in CPython. With it, objects with :meth:`__del__` methods, as well as generators with :keyword:`finally` clauses, can be finalized when they are part of a reference cycle.
As part of this change, module globals are no longer forcibly set to :const:`None` during interpreter shutdown in most cases, instead relying on the normal operation of the cyclic garbage collector. This avoids a whole class of interpreter-shutdown-time errors, usually involving __del__ methods, that have plagued Python since the cyclic GC was first introduced.
PEP 456: Secure and Interchangeable Hash Algorithm
PEP 456 follows up on earlier security fix work done on Python's hash algorithm to address certain DOS attacks to which public facing APIs backed by dictionary lookups may be subject. (See :issue:`14621` for the start of the current round of improvements.) The PEP unifies CPython's hash code to make it easier for a packager to substitute a different hash algorithm, and switches Python's default implementation to a SipHash implementation on platforms that have a 64 bit data type. Any performance differences in comparison with the older FNV algorithm are trivial.
The PEP adds additional fields to the :func:`sys.hash_info` struct sequence to describe the hash algorithm in use by the currently executing binary. Otherwise, the PEP does not alter any existing CPython APIs.
Other build and C API changes
Changes to Python's build process and to the C API include:
- The new :c:func:`Py_SetStandardStreamEncoding` pre-initialization API allows applications embedding the CPython interpreter to reliably force a particular encoding and error handler for the standard streams (Contributed by Bastien Montagne and Nick Coghlan in :issue:`16129`)
- Most Python C APIs that don't mutate string arguments are now correctly marked as accepting const char * rather than char * (Contributed by Serhiy Storchaka in :issue:`1772673`).
"Argument Clinic" (PEP 436) is now part of the CPython build process and can be used to simplify the process of defining and maintaining accurate signatures for builtins and standard library extension modules implemented in C.
The Argument Clinic PEP is not fully up to date with the state of the implementation. This has been deemed acceptable by the release manager and core development team in this case, as Argument Clinic will not be made available as a public API for third party use in Python 3.4.
Unsupported Operating Systems
- Windows 2000
Deprecated Python modules, functions and methods
- :meth:`difflib.SequenceMatcher.isbjunk` and :meth:`difflib.SequenceMatcher.isbpopular` were removed: use x in sm.bjunk and x in sm.bpopular, where sm is a :class:`~difflib.SequenceMatcher` object.
- :func:`importlib.util.module_for_loader` is pending deprecation. Using :func:`importlib.util.module_to_load` and :meth:`importlib.abc.Loader.init_module_attrs` allows subclasses of a loader to more easily customize module loading.
- The :mod:`imp` module is pending deprecation. To keep compatibility with Python 2/3 code bases, the module's removal is currently not scheduled.
- The :mod:`formatter` module is pending deprecation and is slated for removal in Python 3.6.
- MD5 as default digestmod for :mod:`hmac` is deprecated. Python 3.6 will require an explicit digest name or constructor as digestmod argument.
Deprecated functions and types of the C API
- The PyThreadState.tick_counter field has been removed: its value was meaningless since Python 3.2 ("new GIL").
- The site module adding a "site-python" directory to sys.path, if it exists, is deprecated (:issue:`19375`).
Porting to Python 3.4
Changes in the Python API
This section lists previously described changes and other bugfixes that may require changes to your code.
- The ABCs defined in :mod:`importlib.abc` now either raise the appropriate exception or return a default value instead of raising :exc:`NotImplementedError` blindly. This will only affect code calling :func:`super` and falling through all the way to the ABCs. For compatibility, catch both :exc:`NotImplementedError` or the appropriate exception as needed.
- The module type now initializes the :attr:`__package__` and :attr:`__loader__` attributes to None by default. To determine if these attributes were set in a backwards-compatible fashion, use e.g. getattr(module, '__loader__', None) is not None.
- :meth:`importlib.util.module_for_loader` now sets __loader__ and __package__ unconditionally to properly support reloading. If this is not desired then you will need to set these attributes manually. You can use :func:`importlib.util.module_to_load` for module management.
- Import now resets relevant attributes (e.g. __name__, __loader__, __package__, __file__, __cached__) unconditionally when reloading.
- Frozen packages no longer set __path__ to a list containing the package name but an empty list instead. Determing if a module is a package should be done using hasattr(module, '__path__').
- :func:`py_compile.compile` now raises :exc:`FileExistsError` if the file path it would write to is a symlink or a non-regular file. This is to act as a warning that import will overwrite those files with a regular file regardless of what type of file path they were originally.
- :meth:`importlib.abc.SourceLoader.get_source` no longer raises :exc:`ImportError` when the source code being loaded triggers a :exc:`SyntaxError` or :exc:`UnicodeDecodeError`. As :exc:`ImportError` is meant to be raised only when source code cannot be found but it should, it was felt to be over-reaching/overloading of that meaning when the source code is found but improperly structured. If you were catching ImportError before and wish to continue to ignore syntax or decoding issues, catch all three exceptions now.
- :func:`functools.update_wrapper` and :func:`functools.wraps` now correctly set the __wrapped__ attribute to the function being wrapper, even if that function also had its __wrapped__ attribute set. This means __wrapped__ attributes now correctly link a stack of decorated functions rather than every __wrapped__ attribute in the chain referring to the innermost function. Introspection libraries that assumed the previous behaviour was intentional can use :func:`inspect.unwrap` to access the first function in the chain that has no __wrapped__ attribute.
- :class:`importlib.machinery.PathFinder` now passes on the current working directory to objects in :data:`sys.path_hooks` for the empty string. This results in :data:`sys.path_importer_cache` never containing '', thus iterating through :data:`sys.path_importer_cache` based on :data:`sys.path` will not find all keys. A module's __file__ when imported in the current working directory will also now have an absolute path, including when using -m with the interpreter (this does not influence when the path to a file is specified on the command-line).
Changes in the C API
- :c:func:`PyErr_SetImportError` now sets :exc:`TypeError` when its msg argument is not set. Previously only NULL was returned with no exception set.
- The result of the :c:data:`PyOS_ReadlineFunctionPointer` callback must now be a string allocated by :c:func:`PyMem_RawMalloc` or :c:func:`PyMem_RawRealloc`, or NULL if an error occurred, instead of a string allocated by :c:func:`PyMem_Malloc` or :c:func:`PyMem_Realloc`.
- :c:func:`PyThread_set_key_value` now always set the value. In Python 3.3, the function did nothing if the key already exists (if the current value is a non-NULL pointer).
- The f_tstate (thread state) field of the :c:type:`PyFrameObject` structure has been removed to fix a bug: see :issue:`14432` for the rationale.