RFC: PEP xxx: Python Compatibility Version
Hi, I propose the following PEP to add sys.set_python_compat_version(version) to introduce a partial compatibility with Python 3.8 in the next Python 3.9 version. I also propose this PEP in a pull request: https://github.com/python/peps/pull/1209 (Please avoid using the temporary PEP number until this PR is merged: officially, the PEP has no number yet.) If possible, please try to read the whole PEP before replying. I would prefer to avoid knee-jerk reactions :-) The backward compatibility is complex topic where things are not black or white: it's more a grayscale. IMHO with the incoming end of Python 2 support, it's the right time to propose this PEP! Victor PEP: xxx Title: Python Compatibility Version Author: Victor Stinner <vstinner@python.org> Status: Draft Type: Standards Track Content-Type: text/x-rst Created: 18-Oct-2019 Python-Version: 3.9 Abstract ======== Add ``sys.set_python_compat_version(version)`` to enable partial compatibility with requested Python version. Add ``sys.get_python_compat_version()``. Modify a few functions of the standard library to implement a partial compatibility with Python 3.8. Add ``sys.set_python_min_compat_version(version)`` to deny backward compatibility with Python older than *version*. Add ``-X compat_version=VERSION`` and ``-X min_compat_version=VERSION`` command line options. Add ``PYTHONCOMPATVERSION`` and ``PYTHONCOMPATMINVERSION`` environment variables. Rationale ========= The need to evolve frequently ----------------------------- To remain relevant and useful, Python has to evolve frequently. Some enhancements require incompatible changes. Any incompatible change can break an unknown number of Python projects. Developers can decide to not implement a feature because of that. Users want to get the latest Python version to get new features and better performance. A few incompatible changes prevent them to use their applications on the latest Python version. This PEP proposes to add a partial compatibility with old Python versions as a tradeoff to fit both use cases. The main issue with the migration from Python 2 to Python 3 is not that Python 3 is backward incompatible, but how incompatible changes were introduced. Partial compatibility to minimize the Python maintenance burden --------------------------------------------------------------- While technically it would be possible to provide a full compatibility with old Python versions, this PEP proposes to minimize the number of functions handling backward compatibility to reduce the maintenance burden of the Python project ("CPython"). Each change introducing backport compatibility to a function should be properly discussed to estimate the maintenance cost in the long-term. Backward compatibility code will be dropped at each Python release, on a case by case basis. Each compatibility function can be supported for a different number of Python releases depending on its maintenance cost and the estimated risk (number of broken projects) if it's removed. The maintenance cost does not only come from the code implementing the backward compatibility, but come also from additional tests. Cases excluded from backward compatibility ------------------------------------------ The performance overhead of a compatibility code must be low when ``sys.set_python_compat_version()`` is not called. The C API is out of the scope of this PEP: ``Py_LIMITED_API`` macro and the stable ABI are solving this problem differently, see the `PEP 384: Defining a Stable ABI <https://www.python.org/dev/peps/pep-0384/>`_. Security fixes which break the backward compatibility on purpose will not get a compatibility layer. Security matters more than compatibility. For example, ``http.client.HTTPSConnection`` was modified in Python 3.4.3 to performs all the necessary certificate and hostname checks by default. It was a deliberate change motivated by the `PEP 476: Enabling certificate verification by default for stdlib http clients <https://www.python.org/dev/peps/pep-0476/>`_ (`bpo-22417 <https://bugs.python.org/issue22417>`_). The Python language does not provide backward compatibility. Changes which are not clearly incompatible are not covered by this PEP. For example, Python 3.9 changed the default protocol in the ``pickle`` module to Protocol 4 which was first introduced in Python 3.4. This change is backward compatible up to Python 3.4. There is no need to use the Protocol 3 by default when compatibility with Python 3.8 is requested. New ``DeprecationWarning`` and ``PendingDeprecatingWarning`` warnings of Python 3.9 will not be disabled in Python 3.8 compatibility mode. If a project runs its test suite using ``-Werror`` (treat any warning as an error), these warnings must be fixed, or specific deprecation warnings must be ignored on a case by case basis. Upgrade a project to a newer Python ----------------------------------- Without backward compatibility, all incompatible changes must be fixed at once, which can be a blocker issue. It is even worse when a project is upgraded to a newer Python which is separated by multiple releases from the old Python. Postponing an upgrade only makes things worse: each skipped release adds more incompatible changes. The technical debt is only steadily increasing. With backward compatibility, it becomes possible to upgrade Python increamentally in a project, without having to fix all issues at once. The "all-or-nothing" is a showstopper to port large Python 2 code bases to Python 3. The list of incompatible changes between Python 2 and Python 3 is long, and it's getting longer at each Python 3.x release. Cleaning up Python and DeprecationWarning ----------------------------------------- One of the `Zen of Python (PEP 20) <https://www.python.org/dev/peps/pep-0020/>`_ motto is: There should be one-- and preferably only one --obvious way to do it. When Python evolves, new ways emerge inevitably. ``DeprecationWarning`` are emitted to suggest to use the new way, but many developers ignore these warnings, which are silent by default (except in the ``__main__`` module: see the `PEP 565 <https://www.python.org/dev/peps/pep-0565/>_`). Some developers simply ignore all warnings since they are too many warnings, and so only bother with exceptions when deprecated code is removed. Sometimes, supporting both ways has a minor maintenance cost, but developers prefer to drop the old way to clean up the code. Such kind of change is backward incompatible. Some developers can take the end of the Python 2 support as an opportunity to push even more incompatible changes than usual. Adding backward compatibility as an opt-in prevents to break applications and allows developers to continue to do such cleanup. Redistribute the maintenance burden ----------------------------------- The backward compatibility involves authors of backward incompatible changes more in the upgrade path. Examples of backward compatibility ================================== collections ABC aliases ----------------------- ``collections.abc`` aliases to ABC classes have been removed from the ``collections`` module in Python 3.9, after being deprecated since Python 3.3. For example, ``collections.Mapping`` no longer exists. In Python 3.6, aliases were created in ``collections/__init__.py`` by ``from _collections_abc import *``. In Python 3.7, a ``__getattr__()`` has been added to the ``collections`` module to emit a DeprecationWarning at the first access to an attribute:: def __getattr__(name): # For backwards compatibility, continue to make the collections ABCs # through Python 3.6 available through the collections module. # Note, no new collections ABCs were added in Python 3.7 if name in _collections_abc.__all__: obj = getattr(_collections_abc, name) import warnings warnings.warn("Using or importing the ABCs from 'collections' instead " "of from 'collections.abc' is deprecated since Python 3.3, " "and in 3.9 it will stop working", DeprecationWarning, stacklevel=2) globals()[name] = obj return obj raise AttributeError(f'module {__name__!r} has no attribute {name!r}') Compatibility with Python 3.8 can be restored in Python 3.9 by adding back the ``__getattr__()`` function, but only when backward compatibility is requested:: def __getattr__(name): if (sys.get_python_compat_version() < (3, 9) and name in _collections_abc.__all__): ... raise AttributeError(f'module {__name__!r} has no attribute {name!r}') Deprecated open() "U" mode -------------------------- The "U" mode of ``open()`` is deprecated since Python 3.4 and emits a ``DeprecationWarning``. The `bpo-37330 <https://bugs.python.org/issue37330>`_ proposes to drop this mode: ``open()`` would raise an exception if ``U`` mode is used. This change falls into the "cleanup" category: it is not required to implement a feature. A backward compatibility mode would be trivial to implement and would be welcomed here by users. Specification ============= sys functions ------------- Add 3 functions to the ``sys`` module: * ``sys.set_python_compat_version(version)``: set the Python compatibility version. If it has been called previously, use the minimum of requested versions. Raise an exception if ``sys.set_python_min_compat_version(min_version)`` has been called and ``version < min_version``. *version* must be greater than or equal to ``(3, 0)``. * ``sys.set_python_min_compat_version(min_version)``: set the **minimum** compatibility version. Raise an exception if ``sys.set_python_compat_version(old_version)`` has been called previously and ``old_version < min_version``. *min_version* must be greater than or equal to ``(3, 0)``. * ``sys.get_python_compat_version()``: get the Python compatibility version. Return a ``tuple`` of 3 integers. A *version* must a tuple of 2 or 3 integers. ``(major, minor)`` version is equivalent to ``(major, minor, 0)``. By default, ``sys.get_python_compat_version()`` returns the current Python version. Example to request compatibility with Python 3.8.0:: import collections sys.set_python_compat_version((3, 8)) # collections.Mapping alias, removed from Python 3.9, is available # again, even if collections has been imported before calling # set_python_compat_version(). parent = collections.Mapping Obviously, calling ``sys.set_python_compat_version(version)`` has no effect on code executed before the call. Use ``-X compat_version=VERSION`` command line option or ``PYTHONCOMPATVERSIONVERSION=VERSION`` environment variable to set the compatibility version at Python startup. Command line ------------ Add ``-X compat_version=VERSION`` and ``-X min_compat_version=VERSION`` command line options: call respectivelly ``sys.set_python_compat_version()`` and ``sys.set_python_min_compat_version()``. ``VERSION`` is a version string with 2 or 3 numbers (``major.minor.micro`` or ``major.minor``). For example, ``-X compat_version=3.8`` calls ``sys.set_python_compat_version((3, 8))``. Add ``PYTHONCOMPATVERSIONVERSION=VERSION`` and ``PYTHONCOMPATMINVERSION=VERSION=VERSION`` environment variables: call respectivelly ``sys.set_python_compat_version()`` and ``sys.set_python_min_compat_version()``. ``VERSION`` is a version string with the same format that the command line options. Backwards Compatibility ======================= Introducing ``sys.set_python_compat_version()`` function means that an application will behave differently depending on the compatibility version. Moreover, since the version can be decreased multiple times, the application can behave differently depending on the import order. Python 3.9 with ``sys.set_python_compat_version((3, 8))`` is not fully compatible with Python 3.8: the compatibility is only partial. Security Implications ===================== ``sys.set_python_compat_version()`` must not disable security fixes. Alternatives ============ Provide a workaround for each incompatible change ------------------------------------------------- An application can works around most of the incompatible changes which impacts it. For example, ``collections`` aliases can be added again using:: import collections.abc collections.Mapping = collections.abc.Mapping collections.Sequence = collections.abc.Sequence Handle backward compatibility in the parser ------------------------------------------- The parser is modified to support multiple versions of the Python language (grammar). The current Python parser cannot be easily modified for that. AST and grammar are hardcoded to a single Python version. In Python 3.8, ``compile()`` has an undocumented ``_feature_version`` to not consider ``async`` and ``await`` as keywords. The latest major language backward incompatible change was Python 3.7 which made ``async`` and ``await`` real keywords. It seems like Twisted was the only affected project, and Twisted had a single affected function (it used a parameter called ``async``). Handling backward compatibility in the parser seems quite complex, not only to modify the parser, but also for developers who have to check which version of the Python language is used. from __future__ import python38_syntax ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Add ``pythonXY_syntax`` to the ``__future__`` module. It would enable backward compatibility with Python X.Y syntax, but only for the current file. With this option, there is no need to change ``sys.implementation.cache_tag`` to use a different ``.pyc`` filename, since the parser would always produce the same output for the same input (except of the optimization level). Example:: from __future__ import python35_syntax async = 1 await = 2 Update cache_tag ^^^^^^^^^^^^^^^^ Modify the parser to use ``sys.get_python_compat_version()`` to choose the version of the Python language. ``sys.set_python_compat_version()`` updates ``sys.implementation.cache_tag`` to include the compatibility version without the micro version as a suffix. For example, Python 3.9 uses ``'cpython-39'`` by default, but ``sys.set_python_compat_version((3, 7, 2))`` sets ``cache_tag`` to ``'cpython-39-37'``. Changes of the Python language are now allowed in micro releases. One problem is that ``import asyncio`` is likely to fail if ``sys.set_python_compat_version((3, 6))`` has been called previously. The code of the ``asyncio`` module requires ``async`` and ``await`` to be real keywords (change done in Python 3.7). Another problem is that regular users cannot write ``.pyc`` files into system directories, and so cannot create them on demand. It means that ``.pyc`` optimization cannot be used in the backward compatibility mode. One solution for that is to modify the Python installer and Python package installers to precompile ``.pyc`` files not only for the current Python version, but also for multiple older Python versions (up to Python 3.0?). Each ``.py`` file would have 3n ``.pyc`` files (3 optimization levels), where ``n`` is the number of supported Python versions. For example, it means 6 ``.pyc`` files, instead of 3, to support Python 3.8 and Python 3.9. Temporary moratorium on incompatible changes -------------------------------------------- In 2009, the PEP 3003 "Python Language Moratorium" proposed to a temporary moratorium (suspension) of all changes to the Python language syntax, semantics, and built-ins for Python 3.1 and Python 3.2. In May 2018, during PEP 572 discussions, it was also proposed to slow down Python changes: see the python-dev thread `Slow down... <https://mail.python.org/archives/list/python-dev@python.org/thread/HHKRXOMRJQH75VNM3JMSQIOOU6MIUB24/#PHA35EAPNONZMTOYBINGFR6XXNMCDPFQ>`_ `Barry Warsaw's call on this <https://mail.python.org/archives/list/python-dev@python.org/message/XR7IF2OB3S72KBP3PEQ3IKBOERE4FV2I/>`_: I don’t believe that the way for Python to remain relevant and useful for the next 10 years is to cease all language evolution. Who knows what the computing landscape will look like in 5 years, let alone 10? Something as arbitrary as a 10 year moratorium is (again, IMHO) a death sentence for the language. PEP 387 ------- `PEP 387 -- Backwards Compatibility Policy <https://www.python.org/dev/peps/pep-0387/>`_ proposes a process to make incompatible changes. The main point is the 4th step of the process: See if there's any feedback. Users not involved in the original discussions may comment now after seeing the warning. Perhaps reconsider. PEP 497 ------- `PEP 497 -- A standard mechanism for backward compatibility <https://www.python.org/dev/peps/pep-0497/>`_ proposes different solutions to provide backward compatibility. Except of the ``__past__`` mechanism idea, the PEP 497 does not propose concrete solutions: When an incompatible change to core language syntax or semantics is being made, Python-dev's policy is to prefer and expect that, wherever possible, a mechanism for backward compatibility be considered and provided for future Python versions after the breaking change is adopted by default, in addition to any mechanisms proposed for forward compatibility such as new future_statements. Examples of incompatible changes ================================ Python 3.8 ---------- Examples of Python 3.8 incompatible changes: * (During beta phase) ``PyCode_New()`` required a new parameter: it broke all Cython extensions (all projects distributing precompiled Cython code). This change has been reverted during the 3.8 beta phase and a new ``PyCode_NewWithPosOnlyArgs()`` function was added instead. * ``types.CodeType`` requires an additional mandatory parameter. The ``CodeType.replace()`` function was added to help projects to no longer depend on the exact signature of the ``CodeType`` constructor. * C extensions are no longer linked to libpython. * ``sys.abiflags`` changed from ``'m'`` to an empty string. For example, ``python3.8m`` program is gone. * The C structure ``PyInterpreterState`` was made opaque. * Blender: * https://bugzilla.redhat.com/show_bug.cgi?id=1734980#c6 * https://developer.blender.org/D6038 * XML attribute order: `bpo-34160 <https://bugs.python.org/issue34160>`_. Broken projects: * `coverage <https://bugs.python.org/issue34160#msg329612>`_ * `docutils <https://sourceforge.net/p/docutils/bugs/359/>`_ * `pcs <https://bugzilla.redhat.com/show_bug.cgi?id=1705475>`_ * `python-glyphsLib <https://bugzilla.redhat.com/show_bug.cgi?id=1705391>`_ Backward compatibility cannot be added for all these changes. For example, changes in the C API and in the build system are out of the scope of this PEP. See `What’s New In Python 3.8: API and Feature Removals <https://docs.python.org/dev/whatsnew/3.8.html#api-and-feature-removals>`_ for all changes. See also the `Porting to Python 3.8 <https://docs.python.org/dev/whatsnew/3.8.html#porting-to-python-3-8>`_ section of What’s New In Python 3.8. Python 3.7 ---------- Examples of Python 3.7 incompatible changes: * ``async`` and ``await`` are now reserved keywords. * Several undocumented internal imports were removed. One example is that ``os.errno`` is no longer available; use ``import errno`` directly instead. Note that such undocumented internal imports may be removed any time without notice, even in micro version releases. * Unknown escapes consisting of ``'\'`` and an ASCII letter in replacement templates for ``re.sub()`` were deprecated in Python 3.5, and will now cause an error. * The ``asyncio.windows_utils.socketpair()`` function has been removed: it was an alias to ``socket.socketpair()``. * ``asyncio`` no longer exports the ``selectors`` and ``_overlapped`` modules as ``asyncio.selectors`` and ``asyncio._overlapped``. Replace ``from asyncio import selectors`` with ``import selectors``. * PEP 479 is enabled for all code in Python 3.7, meaning that ``StopIteration`` exceptions raised directly or indirectly in coroutines and generators are transformed into ``RuntimeError`` exceptions. * ``socketserver.ThreadingMixIn.server_close()`` now waits until all non-daemon threads complete. Set the new ``block_on_close`` class attribute to ``False`` to get the pre-3.7 behaviour. * The ``struct.Struct.format`` type is now ``str`` instead of ``bytes``. * ``repr`` for ``datetime.timedelta`` has changed to include the keyword arguments in the output. * ``tracemalloc.Traceback`` frames are now sorted from oldest to most recent to be more consistent with ``traceback``. Adding backward compatibility for most of these changes would be easy. See also the `Porting to Python 3.7 <https://docs.python.org/dev/whatsnew/3.7.html#porting-to-python-3-7>`_ section of What’s New In Python 3.7. Micro releases -------------- Sometimes, incompatible changes are introduced in micro releases (``micro`` in ``major.minor.micro``) to fix bugs or security vulnerabilities. Examples: * Python 3.7.2, ``compileall`` and ``py_compile`` module: the *invalidation_mode* parameter's default value is updated to ``None``; the ``SOURCE_DATE_EPOCH`` environment variable no longer overrides the value of the *invalidation_mode* argument, and determines its default value instead. * Python 3.7.1, ``xml`` modules: the SAX parser no longer processes general external entities by default to increase security by default. * Python 3.5.2, ``os.urandom()``: on Linux, if the ``getrandom()`` syscall blocks (the urandom entropy pool is not initialized yet), fall back on reading ``/dev/urandom``. * Python 3.5.1, ``sys.setrecursionlimit()``: a ``RecursionError`` exception is now raised if the new limit is too low at the current recursion depth. * Python 3.4.4, ``ssl.create_default_context()``: RC4 was dropped from the default cipher string. * Python 3.4.3, ``http.client``: ``HTTPSConnection`` now performs all the necessary certificate and hostname checks by default. * Python 3.4.2, ``email.message``: ``EmailMessage.is_attachment()`` is now a method instead of a property, for consistency with ``Message.is_multipart()``. * Python 3.4.1, ``os.makedirs(name, mode=0o777, exist_ok=False)``: Before Python 3.4.1, if *exist_ok* was ``True`` and the directory existed, ``makedirs()`` would still raise an error if *mode* did not match the mode of the existing directory. Since this behavior was impossible to implement safely, it was removed in Python 3.4.1 (`bpo-21082 <https://bugs.python.org/issue21082>`_). Examples of changes made in micro releases which are not backward incompatible: * ``ssl.OP_NO_TLSv1_3`` constant was added to 2.7.15, 3.6.3 and 3.7.0 for backwards compatibility with OpenSSL 1.0.2. * ``typing.AsyncContextManager`` was added to Python 3.6.2. * The ``zipfile`` module accepts a path-like object since Python 3.6.2. * ``loop.create_future()`` was added to Python 3.5.2 in the ``asyncio`` module. No backward compatibility code is needed for such kind of changes. References ========== Accepted PEPs: * `PEP 5 -- Guidelines for Language Evolution <https://www.python.org/dev/peps/pep-0005/>`_ * `PEP 236 -- Back to the __future__ <https://www.python.org/dev/peps/pep-0236/>`_ * `PEP 411 -- Provisional packages in the Python standard library <https://www.python.org/dev/peps/pep-0411/>`_ * `PEP 3002 -- Procedure for Backwards-Incompatible Changes <https://www.python.org/dev/peps/pep-3002/>`_ Draft PEPs: * `PEP 602 -- Annual Release Cycle for Python <https://www.python.org/dev/peps/pep-0602/>`_ * `PEP 605 -- A rolling feature release stream for CPython <https://www.python.org/dev/peps/pep-0605/>`_ * See also withdrawn `PEP 598 -- Introducing incremental feature releases <https://www.python.org/dev/peps/pep-0598/>`_ Copyright ========= This document is placed in the public domain or under the CC0-1.0-Universal license, whichever is more permissive. .. Local Variables: mode: indented-text indent-tabs-mode: nil sentence-end-double-space: t fill-column: 70 coding: utf-8 End: -- Night gathers, and now my watch begins. It shall not end until my death.
The proposal looks awesome to me. Victor :) It could be out of scope for this proposal. How about adding APIs about compatibility implementation in this proposal? e.g compatibility_python_implementation usage will be sys.compatibility_python_implementation(('Cpython', 'PyPy', 'XXXPython')) We can check the implementation name since platform.python_implementation() is already existed. The reason for this feature is that when the python code is running on alternative python interpreter project, user can not sure whether the alternative interpreter could run as the code author intended. This feature would help them(author, user). 2019년 10월 18일 (금) 오전 10:19, Victor Stinner <vstinner@python.org>님이 작성:
-- Software Development Engineer at Kakao corp. Tel: +82 010-3353-9127 Email: donghee.na92@gmail.com | denny.i@kakaocorp.com Linkedin: https://www.linkedin.com/in/dong-hee-na-2b713b49/
On 10/17/2019 06:02 PM, Victor Stinner wrote:
-1
If possible, please try to read the whole PEP before replying.
Done. TL;DR Besides being a maintenance burden, and one more source of bugs, it will be a debugging nightmare -- especially if different levels of "back-compatibility" have been selected.
Is "developer" an app developer or a library developer? Either way, such developers are constrained by the lowest python version they support. How many versions back are we going to allow? If we don't support each and every breaking change then some application/library is going to be inoperable at the latest Python version, even with "back compatibility".
If they have to enable back-compatibility mode they still aren't getting the latest features/performance. Also, if the app isn't using these new features (even without back-compatibility), how are they benefiting the user?
How will we estimate this number? What's the minimum? Is that minimum number dependent on the user base of those projects? How do we estimate that?
Because the technical debt is now burdening Python instead.
Are you saying we are going to add a Python2.7 compatibility layer? -- ~Ethan~
Hi Ethan, Le ven. 18 oct. 2019 à 04:38, Ethan Furman <ethan@stoneleaf.us> a écrit :
I hesitated to use "core developers" instead of "developers" in the PEP, since CPython is not only maintained by "core developers". But it seems like it introduced some confusion. This paragraph is about CPython changes and core developers. Core developers love to drop legacy code :-)
How many versions back are we going to allow?
I'm not sure that I correctly understand your question. Are you asking if I plan to only support Python 3.8, or if I plan to add support for Python 3.7 and older, in Python 3.9? If it is your question: I propose to add a partial Python 3.8 support in 3.9, then extend this backward compatiblity with Python 3.9 in Python 3.10, etc. If you are asking how many releases a backward compatibility will be supported, I would say that it should be discussed on a case by case basis. For example, let's say that Python 3.9 provides backward compatibility for collections ABC aliases and open() "U" mode. We may remove collections ABC aliases in Python 3.10, but keep open() "U" mode in Python 3.10 and only remove it from Python 3.11. Well, it's not different from what we already do with PendingDeprecationWarning, DeprecationWarning and the final feature removal. The deprecation duration is different for each deprecated function.
If we don't support each and every breaking change then some application/library is going to be inoperable at the latest Python version, even with "back compatibility".
Right, the PEP doesn't solve all compatibility issues. It proposes a compromise between "break all applications" (no bw compat) and "fix all applications" (full bw compat). It's a tradeoff depending on the CPython maintenance burden.
If they have to enable back-compatibility mode they still aren't getting the latest features/performance.
Why not? If you run your application on Python 3.9 and Python 3.9 is faster, your app will run faster, no? About latest features. Usually, Python modules support a range of Python versions, and are able to use newer functions when available. Let's say that MyApp application uses modules A and B. A is fully compatible with Python 3.9 and so is able to use latest cool Python features. The module B wasn't ported yet and so requests Python 3.8 backward compatibility. This doesn't remove Python 3.9 new features: they remain available.
Also, if the app isn't using these new features (even without back-compatibility), how are they benefiting the user?
If you ignore new features and better performance, there is still an advantage: being to run an old application unmodified, or at least to reduce the time needed to port it. Let's say that you have an old application which doesn't work unmodified on Python 3.9, but your system only provides Python 3.9. How do you run the application? Attempt "python3.9 -X compat_version=3.9 app.py". If you are lucky, the backward compatibility is enough to run the application unmodified. Great, isn't it? If not, at least the backward compatibility should reduce the number of changes needed to run the application.
I have no answer to this question. This problem isn't different from the current problem of introducing incompatible changes in Python on purpose. We can just hope that few projects will break. Or we still have the option of reverting the change if we were not lucky this time. It's all about luck currently :-) Right now, I have no idea how many projects are broken by the removal of collections ABC aliases. I'm confident that it's greater than 0, since we have to wait until pip is updated to stop using the deprecated alises. Even pip ignores DeprecationWarning... (In practice, it was a dependency which has been fixed long time ago, but there was no release...)
It's exactly the purpose of the PEP ;-) I tried to make it obvious in the "Redistribute the maintenance burden" section. If core developers want to break Python, fine! But they should help somehow to update projects written in Python, at least the most popular Python modules on PyPI. It's already the case, most of fixes to support the latest Python version are done by core developers.
I'm not sure if it's a real question or not. I only plan to add the compatibility with Python 3.8 and newer. Right now, I don't want to propose how many releases the compatibility of a specific function will be supported. If the PEP is accepted and implemented, I would prefer to see how it works in practice. Victor -- Night gathers, and now my watch begins. It shall not end until my death.
On Thu, Oct 17, 2019 at 7:39 PM Ethan Furman <ethan@stoneleaf.us> wrote:
-1 from me as well.
That was my immediate thought as well as you now have to test that code works in the stdlib for all potential version values to make sure all side-effects/ramifications are accounted for. And you already have sys.version_info for making sure you have a minimum version support for your code. Basically if people can't update yet then they simply can't update yet. Staying on an older version that they were already happy enough with to support shouldn't lead to a maintenance burden just to try and give them free performance perks at a cost to the core devs. -Brett
I re-read your proposal. Please ignore my suggestion. :) I misunderstood core concept of your proposal :) 2019년 10월 18일 (금) 오전 10:19, Victor Stinner <vstinner@python.org>님이 작성:
-- Software Development Engineer at Kakao corp. Tel: +82 010-3353-9127 Email: donghee.na92@gmail.com | denny.i@kakaocorp.com Linkedin: https://www.linkedin.com/in/dong-hee-na-2b713b49/
On 2019-10-17 18:02, Victor Stinner wrote:
I read the PEP, but I find it a bit difficult to understand what it's actually proposing. It seems like you are proposing that there be this function set_python_compat_version which will effectively act as a global flag affecting various functionality in the standard library, namely by making it work as it did in a past version of Python. Is that the basic idea? My main point of confusion is that I don't see an explicit list of what functionality will be supported by this flag. In the abstract it says "Modify a few functions of the standard library to implement a partial compatibility". What precisely are those functions? Because of this omission, I can't tell if this PEP is actually proposing a specific set of accommodations to backwards compatibility, or just proposing a mechanism for handling such accommodations. If the PEP is proposing that existing stdlib functions be modified to actually check the compatibility version and give different behavior accordingly, I think there needs to be an exhaustive list of which functions those will be and what exactly the flag will do. If it's just proposing a mechanism, without necessarily proposing that any actual changes be made to use that mechanism, okay, but I'm somewhat skeptical that that will get much traction, Actually I'm somewhat skeptical it'll get much traction anyway. The overall issue seems to be more a human issue of governance and priorities. The problem is, even supposing this mechanism is in place, who is going to enforce it to ensure that any time a backwards-incompatible change is made, it is accompanied by work to make the changes "undoable" via this "compat version" mechanism? Having a function called set_python_compat_version is no good unless, going forward, every backwards-incompatible change is blockable this way, because otherwise someone who relies on old behavior that isn't "compatibilized" will still be unable to upgrade in the same way you describe in your rationale. In other words, it doesn't matter if we have set_python_compat_version, what matters is how fully it actually covers the backwards compatible changes. It seems like enforcing a policy of "every backwards compatible change must be compatabilizable in this way" is more a matter for the dev team's internal guidelines (akin to "you must provide unit tests") than for a PEP. The part that confuses me the most is this: "Backward compatibility code will be dropped at each Python release, on a case by case basis. Each compatibility function can be supported for a different number of Python releases depending on its maintenance cost and the estimated risk (number of broken projects) if it's removed." Isn't that sort of already what deprecation schedules and the like are supposed to accomplish? It seems like this paragraph totally undercuts the whole point of the proposal. Right now maybe we have a situation where if I'm on Python 4.8 and the newest version is 4.9, I could be bitten by backwards incompatibility. With this change, I could be on Python 4.8 and maybe the newest is 4.12 and I'm hanging on by setting compat mode. But if the specific old features I haven't yet phased out are the ones dropped from compat mode in 4.13, then the whole mechanism is useless to me. So it just delays the incompatibility barrier a few versions, which is what deprecation schedules are already supposed to do. If backwards compatibility is itself going to be backwards-incompatibly removed, how is it better to use this system than to just give every breaking change a deprecation schedule that will phase it out over the same amount of time? I'm leaving aside some of the even hairier corner cases, such as the situation mentioned in the PEP where behavior could depend on import order because the behavior of compatibilized functions may change on the fly if multiple calls to set_compat_version occur. That sounds like a recipe for disaster, and suggests to me that the whole issue is better handled with some kind of external declarative system (akin to pip or conda dependency tracking) rather than a Python-internal system which is subject to the vagaries of execution order. As I said, I may be misunderstanding the PEP, but either way I think it could benefit from some editing, because the way it is structured right now is (to me) quite confusing. It just starts off by saying it will "add sys.set_python_compat_version(version) to enable partial compatibility" and then claims it will "modify a few functions of the standard library" to implement this, but I can't see anywhere that it says what those functions are, and because of that I'm no longer even sure what it means by "enable partial incompatibility". I think the "specification" section needs to be a lot more, well, specific. Also, just as a minor point, I don't think it is a good idea to name the function "set_python_compat_version" if it's not actually setting the version --- which it isn't because according to the PEP the actual compat version will be set to the minimum of any versions passed if this function is called multiple times. That's not what "setting" the version means to me; if I set something to 2 and then I set it to 3, I expect its value to then be 3, but with this it would still be 2. A better choice would be a name like "require_compatibility_with", which is more compatible (har har) with the idea that multiple calls coalesce into their minimum. Likewise set_python_min_compat_version might be better as "prohibit_compatibility_with" or something. -- Brendan Barnwell "Do not follow where the path may lead. Go, instead, where there is no path, and leave a trail." --author unknown
Hi Victor, I think this is a nice idea, but I do wonder how this would be used in practice: The problem is that Python projects typically use lots of 3rd party modules and if a few decide to use/rely on the feature, this may create a situation where packages then compete over which compatibility mode to enable. In such situations, I typically prefer failing early, since detecting whether a package requires an older implementation of a Python API is hard from the outside. Regardless of the outcome of the PEP, what I think would be a great addition is a place where to find the the list of incompatible changes per Python release in a single document - much like you have done in the PEP. Cheers, -- Marc-Andre Lemburg eGenix.com Professional Python Services directly from the Experts
::: We implement business ideas - efficiently in both time and costs ::: eGenix.com Software, Skills and Services GmbH Pastor-Loeh-Str.48 D-40764 Langenfeld, Germany. CEO Dipl.-Math. Marc-Andre Lemburg Registered at Amtsgericht Duesseldorf: HRB 46611 http://www.egenix.com/company/contact/ http://www.malemburg.com/ On 18.10.2019 03:02, Victor Stinner wrote:
On Fri, Oct 18, 2019 at 3:18 AM Victor Stinner <vstinner@python.org> wrote:
The proposal seems to add another dimension to the complexity the language is already facing. So instead of trying to keep up with different versions (and how they affect the core functionality or stdlib, or 3rd party libs), it adds another degree of freedom to area, so it is no longer identified by the version it supports only, but also by (possibly) a list of compatibility versions it also supports and on which functionality subset. Determining the target version for the project in this environment will become a serious task for an optimizer: If someone maintains a project on version A and version B introduces some new features and breaks some other things he may: 1) either backport the relevant new things from B to his project, 2) migrate the project from A to B 3) leave it be Whatever he will choose would depend on the importance and effort required and I would consider this to be a "standard" approach and I do not see anything wrong with that. Now, let's say B introduces some new features and breaks some things and provides B(a) compatibility mode for 'a' functionality. Now the project maintainer has new options (apart from the previous ones): 4) Convince Python devs to define 'a' as such that B(a) will cover his project "legacy" code, so he would be able to switch to B(a) without doing anything. 5) If the lobbying does not succeed he still may be able to use B(a) and just migrate from A to B(a) (assuming that A->B(a) is less effort than full A->B). In both last cases the project will find itself in a "compatibilty version" in between, i.e. neither runs on pure A, nor on pure B. Now imagine the project will use 3rd party lib, which may run on A, or B, but would not support B(a) (i.e. mixing the different features from both versions A and B will break the lib). Maybe he will be able to run the lib in a restricted "pure A" version mode (because the project will not need the other non-compatible functionality), or in a restricted "pure B" (for the same reason), or he may try to convince the 3rd party lib dev to support also B(a). Would I venture into B(a) land with my project? Definitely not. But I am writing "small stuff" (experimental code, which usually depends on 3rd party code as well). I would be interested to know, from people who do the "serious stuff", business logic, production code, how they feel about it and how they plan to do lobbying at the core-devs to get the right B(a). Richard
Hi, Just bear in mind that I have read only a couple sentences in the PEP. But as soon as you propose to offer only "partial compatibility" with a previous version, then I think it's worse than useless. You are introducing an additional Python version, i.e. "3.9 bended towards 3.8 partial compatibility", which is neither fully 3.8 nor fully 3.9. It won't make life any easier for third-party library maintainers, on the contrary: it will make it worse. Regards Antoine. On Fri, 18 Oct 2019 03:02:50 +0200 Victor Stinner <vstinner@python.org> wrote:
The proposal looks awesome to me. Victor :) It could be out of scope for this proposal. How about adding APIs about compatibility implementation in this proposal? e.g compatibility_python_implementation usage will be sys.compatibility_python_implementation(('Cpython', 'PyPy', 'XXXPython')) We can check the implementation name since platform.python_implementation() is already existed. The reason for this feature is that when the python code is running on alternative python interpreter project, user can not sure whether the alternative interpreter could run as the code author intended. This feature would help them(author, user). 2019년 10월 18일 (금) 오전 10:19, Victor Stinner <vstinner@python.org>님이 작성:
-- Software Development Engineer at Kakao corp. Tel: +82 010-3353-9127 Email: donghee.na92@gmail.com | denny.i@kakaocorp.com Linkedin: https://www.linkedin.com/in/dong-hee-na-2b713b49/
On 10/17/2019 06:02 PM, Victor Stinner wrote:
-1
If possible, please try to read the whole PEP before replying.
Done. TL;DR Besides being a maintenance burden, and one more source of bugs, it will be a debugging nightmare -- especially if different levels of "back-compatibility" have been selected.
Is "developer" an app developer or a library developer? Either way, such developers are constrained by the lowest python version they support. How many versions back are we going to allow? If we don't support each and every breaking change then some application/library is going to be inoperable at the latest Python version, even with "back compatibility".
If they have to enable back-compatibility mode they still aren't getting the latest features/performance. Also, if the app isn't using these new features (even without back-compatibility), how are they benefiting the user?
How will we estimate this number? What's the minimum? Is that minimum number dependent on the user base of those projects? How do we estimate that?
Because the technical debt is now burdening Python instead.
Are you saying we are going to add a Python2.7 compatibility layer? -- ~Ethan~
Hi Ethan, Le ven. 18 oct. 2019 à 04:38, Ethan Furman <ethan@stoneleaf.us> a écrit :
I hesitated to use "core developers" instead of "developers" in the PEP, since CPython is not only maintained by "core developers". But it seems like it introduced some confusion. This paragraph is about CPython changes and core developers. Core developers love to drop legacy code :-)
How many versions back are we going to allow?
I'm not sure that I correctly understand your question. Are you asking if I plan to only support Python 3.8, or if I plan to add support for Python 3.7 and older, in Python 3.9? If it is your question: I propose to add a partial Python 3.8 support in 3.9, then extend this backward compatiblity with Python 3.9 in Python 3.10, etc. If you are asking how many releases a backward compatibility will be supported, I would say that it should be discussed on a case by case basis. For example, let's say that Python 3.9 provides backward compatibility for collections ABC aliases and open() "U" mode. We may remove collections ABC aliases in Python 3.10, but keep open() "U" mode in Python 3.10 and only remove it from Python 3.11. Well, it's not different from what we already do with PendingDeprecationWarning, DeprecationWarning and the final feature removal. The deprecation duration is different for each deprecated function.
If we don't support each and every breaking change then some application/library is going to be inoperable at the latest Python version, even with "back compatibility".
Right, the PEP doesn't solve all compatibility issues. It proposes a compromise between "break all applications" (no bw compat) and "fix all applications" (full bw compat). It's a tradeoff depending on the CPython maintenance burden.
If they have to enable back-compatibility mode they still aren't getting the latest features/performance.
Why not? If you run your application on Python 3.9 and Python 3.9 is faster, your app will run faster, no? About latest features. Usually, Python modules support a range of Python versions, and are able to use newer functions when available. Let's say that MyApp application uses modules A and B. A is fully compatible with Python 3.9 and so is able to use latest cool Python features. The module B wasn't ported yet and so requests Python 3.8 backward compatibility. This doesn't remove Python 3.9 new features: they remain available.
Also, if the app isn't using these new features (even without back-compatibility), how are they benefiting the user?
If you ignore new features and better performance, there is still an advantage: being to run an old application unmodified, or at least to reduce the time needed to port it. Let's say that you have an old application which doesn't work unmodified on Python 3.9, but your system only provides Python 3.9. How do you run the application? Attempt "python3.9 -X compat_version=3.9 app.py". If you are lucky, the backward compatibility is enough to run the application unmodified. Great, isn't it? If not, at least the backward compatibility should reduce the number of changes needed to run the application.
I have no answer to this question. This problem isn't different from the current problem of introducing incompatible changes in Python on purpose. We can just hope that few projects will break. Or we still have the option of reverting the change if we were not lucky this time. It's all about luck currently :-) Right now, I have no idea how many projects are broken by the removal of collections ABC aliases. I'm confident that it's greater than 0, since we have to wait until pip is updated to stop using the deprecated alises. Even pip ignores DeprecationWarning... (In practice, it was a dependency which has been fixed long time ago, but there was no release...)
It's exactly the purpose of the PEP ;-) I tried to make it obvious in the "Redistribute the maintenance burden" section. If core developers want to break Python, fine! But they should help somehow to update projects written in Python, at least the most popular Python modules on PyPI. It's already the case, most of fixes to support the latest Python version are done by core developers.
I'm not sure if it's a real question or not. I only plan to add the compatibility with Python 3.8 and newer. Right now, I don't want to propose how many releases the compatibility of a specific function will be supported. If the PEP is accepted and implemented, I would prefer to see how it works in practice. Victor -- Night gathers, and now my watch begins. It shall not end until my death.
On Thu, Oct 17, 2019 at 7:39 PM Ethan Furman <ethan@stoneleaf.us> wrote:
-1 from me as well.
That was my immediate thought as well as you now have to test that code works in the stdlib for all potential version values to make sure all side-effects/ramifications are accounted for. And you already have sys.version_info for making sure you have a minimum version support for your code. Basically if people can't update yet then they simply can't update yet. Staying on an older version that they were already happy enough with to support shouldn't lead to a maintenance burden just to try and give them free performance perks at a cost to the core devs. -Brett
I re-read your proposal. Please ignore my suggestion. :) I misunderstood core concept of your proposal :) 2019년 10월 18일 (금) 오전 10:19, Victor Stinner <vstinner@python.org>님이 작성:
-- Software Development Engineer at Kakao corp. Tel: +82 010-3353-9127 Email: donghee.na92@gmail.com | denny.i@kakaocorp.com Linkedin: https://www.linkedin.com/in/dong-hee-na-2b713b49/
On 2019-10-17 18:02, Victor Stinner wrote:
I read the PEP, but I find it a bit difficult to understand what it's actually proposing. It seems like you are proposing that there be this function set_python_compat_version which will effectively act as a global flag affecting various functionality in the standard library, namely by making it work as it did in a past version of Python. Is that the basic idea? My main point of confusion is that I don't see an explicit list of what functionality will be supported by this flag. In the abstract it says "Modify a few functions of the standard library to implement a partial compatibility". What precisely are those functions? Because of this omission, I can't tell if this PEP is actually proposing a specific set of accommodations to backwards compatibility, or just proposing a mechanism for handling such accommodations. If the PEP is proposing that existing stdlib functions be modified to actually check the compatibility version and give different behavior accordingly, I think there needs to be an exhaustive list of which functions those will be and what exactly the flag will do. If it's just proposing a mechanism, without necessarily proposing that any actual changes be made to use that mechanism, okay, but I'm somewhat skeptical that that will get much traction, Actually I'm somewhat skeptical it'll get much traction anyway. The overall issue seems to be more a human issue of governance and priorities. The problem is, even supposing this mechanism is in place, who is going to enforce it to ensure that any time a backwards-incompatible change is made, it is accompanied by work to make the changes "undoable" via this "compat version" mechanism? Having a function called set_python_compat_version is no good unless, going forward, every backwards-incompatible change is blockable this way, because otherwise someone who relies on old behavior that isn't "compatibilized" will still be unable to upgrade in the same way you describe in your rationale. In other words, it doesn't matter if we have set_python_compat_version, what matters is how fully it actually covers the backwards compatible changes. It seems like enforcing a policy of "every backwards compatible change must be compatabilizable in this way" is more a matter for the dev team's internal guidelines (akin to "you must provide unit tests") than for a PEP. The part that confuses me the most is this: "Backward compatibility code will be dropped at each Python release, on a case by case basis. Each compatibility function can be supported for a different number of Python releases depending on its maintenance cost and the estimated risk (number of broken projects) if it's removed." Isn't that sort of already what deprecation schedules and the like are supposed to accomplish? It seems like this paragraph totally undercuts the whole point of the proposal. Right now maybe we have a situation where if I'm on Python 4.8 and the newest version is 4.9, I could be bitten by backwards incompatibility. With this change, I could be on Python 4.8 and maybe the newest is 4.12 and I'm hanging on by setting compat mode. But if the specific old features I haven't yet phased out are the ones dropped from compat mode in 4.13, then the whole mechanism is useless to me. So it just delays the incompatibility barrier a few versions, which is what deprecation schedules are already supposed to do. If backwards compatibility is itself going to be backwards-incompatibly removed, how is it better to use this system than to just give every breaking change a deprecation schedule that will phase it out over the same amount of time? I'm leaving aside some of the even hairier corner cases, such as the situation mentioned in the PEP where behavior could depend on import order because the behavior of compatibilized functions may change on the fly if multiple calls to set_compat_version occur. That sounds like a recipe for disaster, and suggests to me that the whole issue is better handled with some kind of external declarative system (akin to pip or conda dependency tracking) rather than a Python-internal system which is subject to the vagaries of execution order. As I said, I may be misunderstanding the PEP, but either way I think it could benefit from some editing, because the way it is structured right now is (to me) quite confusing. It just starts off by saying it will "add sys.set_python_compat_version(version) to enable partial compatibility" and then claims it will "modify a few functions of the standard library" to implement this, but I can't see anywhere that it says what those functions are, and because of that I'm no longer even sure what it means by "enable partial incompatibility". I think the "specification" section needs to be a lot more, well, specific. Also, just as a minor point, I don't think it is a good idea to name the function "set_python_compat_version" if it's not actually setting the version --- which it isn't because according to the PEP the actual compat version will be set to the minimum of any versions passed if this function is called multiple times. That's not what "setting" the version means to me; if I set something to 2 and then I set it to 3, I expect its value to then be 3, but with this it would still be 2. A better choice would be a name like "require_compatibility_with", which is more compatible (har har) with the idea that multiple calls coalesce into their minimum. Likewise set_python_min_compat_version might be better as "prohibit_compatibility_with" or something. -- Brendan Barnwell "Do not follow where the path may lead. Go, instead, where there is no path, and leave a trail." --author unknown
Hi Victor, I think this is a nice idea, but I do wonder how this would be used in practice: The problem is that Python projects typically use lots of 3rd party modules and if a few decide to use/rely on the feature, this may create a situation where packages then compete over which compatibility mode to enable. In such situations, I typically prefer failing early, since detecting whether a package requires an older implementation of a Python API is hard from the outside. Regardless of the outcome of the PEP, what I think would be a great addition is a place where to find the the list of incompatible changes per Python release in a single document - much like you have done in the PEP. Cheers, -- Marc-Andre Lemburg eGenix.com Professional Python Services directly from the Experts
::: We implement business ideas - efficiently in both time and costs ::: eGenix.com Software, Skills and Services GmbH Pastor-Loeh-Str.48 D-40764 Langenfeld, Germany. CEO Dipl.-Math. Marc-Andre Lemburg Registered at Amtsgericht Duesseldorf: HRB 46611 http://www.egenix.com/company/contact/ http://www.malemburg.com/ On 18.10.2019 03:02, Victor Stinner wrote:
On Fri, Oct 18, 2019 at 3:18 AM Victor Stinner <vstinner@python.org> wrote:
The proposal seems to add another dimension to the complexity the language is already facing. So instead of trying to keep up with different versions (and how they affect the core functionality or stdlib, or 3rd party libs), it adds another degree of freedom to area, so it is no longer identified by the version it supports only, but also by (possibly) a list of compatibility versions it also supports and on which functionality subset. Determining the target version for the project in this environment will become a serious task for an optimizer: If someone maintains a project on version A and version B introduces some new features and breaks some other things he may: 1) either backport the relevant new things from B to his project, 2) migrate the project from A to B 3) leave it be Whatever he will choose would depend on the importance and effort required and I would consider this to be a "standard" approach and I do not see anything wrong with that. Now, let's say B introduces some new features and breaks some things and provides B(a) compatibility mode for 'a' functionality. Now the project maintainer has new options (apart from the previous ones): 4) Convince Python devs to define 'a' as such that B(a) will cover his project "legacy" code, so he would be able to switch to B(a) without doing anything. 5) If the lobbying does not succeed he still may be able to use B(a) and just migrate from A to B(a) (assuming that A->B(a) is less effort than full A->B). In both last cases the project will find itself in a "compatibilty version" in between, i.e. neither runs on pure A, nor on pure B. Now imagine the project will use 3rd party lib, which may run on A, or B, but would not support B(a) (i.e. mixing the different features from both versions A and B will break the lib). Maybe he will be able to run the lib in a restricted "pure A" version mode (because the project will not need the other non-compatible functionality), or in a restricted "pure B" (for the same reason), or he may try to convince the 3rd party lib dev to support also B(a). Would I venture into B(a) land with my project? Definitely not. But I am writing "small stuff" (experimental code, which usually depends on 3rd party code as well). I would be interested to know, from people who do the "serious stuff", business logic, production code, how they feel about it and how they plan to do lobbying at the core-devs to get the right B(a). Richard
Hi, Just bear in mind that I have read only a couple sentences in the PEP. But as soon as you propose to offer only "partial compatibility" with a previous version, then I think it's worse than useless. You are introducing an additional Python version, i.e. "3.9 bended towards 3.8 partial compatibility", which is neither fully 3.8 nor fully 3.9. It won't make life any easier for third-party library maintainers, on the contrary: it will make it worse. Regards Antoine. On Fri, 18 Oct 2019 03:02:50 +0200 Victor Stinner <vstinner@python.org> wrote:
participants (8)
-
Antoine Pitrou -
Brendan Barnwell -
Brett Cannon -
Dong-hee Na -
Ethan Furman -
M.-A. Lemburg -
Richard Musil -
Victor Stinner