python-peps/peps/pep-0413.rst

923 lines
37 KiB
ReStructuredText

PEP: 413
Title: Faster evolution of the Python Standard Library
Version: $Revision$
Last-Modified: $Date$
Author: Alyssa Coghlan <ncoghlan@gmail.com>
Status: Withdrawn
Type: Process
Content-Type: text/x-rst
Created: 24-Feb-2012
Post-History: 24-Feb-2012, 25-Feb-2012
PEP Withdrawal
==============
With the acceptance of :pep:`453` meaning that ``pip`` will be available to
most new Python users by default, this will hopefully reduce the pressure
to add new modules to the standard library before they are sufficiently
mature.
The last couple of years have also seen increased usage of the model where
a standard library package also has an equivalent available from the Python
Package Index that also supports older versions of Python.
Given these two developments and the level of engagement throughout the
Python 3.4 release cycle, the PEP author no longer feels it would be
appropriate to make such a fundamental change to the standard library
development process.
Abstract
========
This PEP proposes the adoption of a separate versioning scheme for the
standard library (distinct from, but coupled to, the existing language
versioning scheme) that allows accelerated releases of the Python standard
library, while maintaining (or even slowing down) the current rate of
change in the core language definition.
Like :pep:`407`, it aims to adjust the current balance between measured
change that allows the broader community time to adapt and being able to
keep pace with external influences that evolve more rapidly than the current
release cycle can handle (this problem is particularly notable for
standard library elements that relate to web technologies).
However, it's more conservative in its aims than :pep:`407`, seeking to
restrict the increased pace of development to builtin and standard library
interfaces, without affecting the rate of change for other elements such
as the language syntax and version numbering as well as the CPython
binary API and bytecode format.
Rationale
=========
To quote the :pep:`407` abstract:
Finding a release cycle for an open-source project is a delicate exercise
in managing mutually contradicting constraints: developer manpower,
availability of release management volunteers, ease of maintenance for
users and third-party packagers, quick availability of new features (and
behavioural changes), availability of bug fixes without pulling in new
features or behavioural changes.
The current release cycle errs on the conservative side. It is adequate
for people who value stability over reactivity. This PEP is an attempt to
keep the stability that has become a Python trademark, while offering a
more fluid release of features, by introducing the notion of long-term
support versions.
I agree with the :pep:`407` authors that the current release cycle of the
*standard library* is too slow to effectively cope with the pace of change
in some key programming areas (specifically, web protocols and related
technologies, including databases, templating and serialisation formats).
However, I have written this competing PEP because I believe that the
approach proposed in :pep:`407` of offering full, potentially binary
incompatible releases of CPython every 6 months places too great a burden
on the wider Python ecosystem.
Under the current CPython release cycle, distributors of key binary
extensions will often support Python releases even after the CPython branches
enter "security fix only" mode (for example, Twisted currently ships binaries
for 2.5, 2.6 and 2.7, NumPy and SciPy support those 3 along with 3.1 and 3.2,
PyGame adds a 2.4 binary release, wxPython provides both 32-bit and 64-bit
binaries for 2.6 and 2.7, etc).
If CPython were to triple (or more) its rate of releases, the developers of
those libraries (many of which are even more resource starved than CPython)
would face an unpalatable choice: either adopt the faster release cycle
themselves (up to 18 simultaneous binary releases for PyGame!), drop
older Python versions more quickly, or else tell their users to stick to the
CPython LTS releases (thus defeating the entire point of speeding up the
CPython release cycle in the first place).
Similarly, many support tools for Python (e.g. syntax highlighters) can take
quite some time to catch up with language level changes.
At a cultural level, the Python community is also accustomed to a certain
meaning for Python version numbers - they're linked to deprecation periods,
support periods, all sorts of things. :pep:`407` proposes that collective
knowledge all be swept aside, without offering a compelling rationale for why
such a course of action is actually *necessary* (aside from, perhaps, making
the lives of the CPython core developers a little easier at the expense of
everyone else).
However, if we go back to the primary rationale for increasing the pace of
change (i.e. more timely support for web protocols and related technologies),
we can note that those only require *standard library* changes. That means
many (perhaps even most) of the negative effects on the wider community can
be avoided by explicitly limiting which parts of CPython are affected by the
new release cycle, and allowing other parts to evolve at their current, more
sedate, pace.
Proposal
========
This PEP proposes the introduction of a new kind of CPython release:
"standard library releases". As with :pep:`407`, this will give CPython 3 kinds
of release:
* Language release: "x.y.0"
* Maintenance release: "x.y.z" (where z > 0)
* Standard library release: "x.y (xy.z)" (where z > 0)
Under this scheme, an unqualified version reference (such as "3.3") would
always refer to the most recent corresponding language or maintenance
release. It will never be used without qualification to refer to a standard
library release (at least, not by python-dev - obviously, we can only set an
example, not force the rest of the Python ecosystem to go along with it).
Language releases will continue as they are now, as new versions of the
Python language definition, along with a new version of the CPython
interpreter and the Python standard library. Accordingly, a language
release may contain any and all of the following changes:
* new language syntax
* new standard library changes (see below)
* new deprecation warnings
* removal of previously deprecated features
* changes to the emitted bytecode
* changes to the AST
* any other significant changes to the compilation toolchain
* changes to the core interpreter eval loop
* binary incompatible changes to the C ABI (although the :pep:`384` stable ABI
must still be preserved)
* bug fixes
Maintenance releases will also continue as they do today, being strictly
limited to bug fixes for the corresponding language release. No new features
or radical internal changes are permitted.
The new standard library releases will occur in parallel with each
maintenance release and will be qualified with a new version identifier
documenting the standard library version. Standard library releases may
include the following changes:
* new features in pure Python modules
* new features in C extension modules (subject to :pep:`399` compatibility
requirements)
* new features in language builtins (provided the C ABI remains unaffected)
* bug fixes from the corresponding maintenance release
Standard library version identifiers are constructed by combining the major
and minor version numbers for the Python language release into a single two
digit number and then appending a sequential standard library version
identifier.
Release Cycle
-------------
When maintenance releases are created, *two* new versions of Python would
actually be published on python.org (using the first 3.3 maintenance release,
planned for February 2013 as an example)::
3.3.1 # Maintenance release
3.3 (33.1) # Standard library release
A further 6 months later, the next 3.3 maintenance release would again be
accompanied by a new standard library release::
3.3.2 # Maintenance release
3.3 (33.2) # Standard library release
Again, the standard library release would be binary compatible with the
previous language release, merely offering additional features at the
Python level.
Finally, 18 months after the release of 3.3, a new language release would
be made around the same time as the final 3.3 maintenance and standard
library releases::
3.3.3 # Maintenance release
3.3 (33.3) # Standard library release
3.4.0 # Language release
The 3.4 release cycle would then follow a similar pattern to that for 3.3::
3.4.1 # Maintenance release
3.4 (34.1) # Standard library release
3.4.2 # Maintenance release
3.4 (34.2) # Standard library release
3.4.3 # Maintenance release
3.4 (34.3) # Standard library release
3.5.0 # Language release
Programmatic Version Identification
-----------------------------------
To expose the new version details programmatically, this PEP proposes the
addition of a new ``sys.stdlib_info`` attribute that records the new
standard library version above and beyond the underlying interpreter
version. Using the initial Python 3.3 release as an example::
sys.stdlib_info(python=33, version=0, releaselevel='final', serial=0)
This information would also be included in the ``sys.version`` string::
Python 3.3.0 (33.0, default, Feb 17 2012, 23:03:41)
[GCC 4.6.1]
Security Fixes and Other "Out of Cycle" Releases
------------------------------------------------
For maintenance releases the process of handling out-of-cycle releases (for
example, to fix a security issue or resolve a critical bug in a new release),
remains the same as it is now: the minor version number is incremented and a
new release is made incorporating the required bug fixes, as well as any
other bug fixes that have been committed since the previous release.
For standard library releases, the process is essentially the same, but the
corresponding "What's New?" document may require some tidying up for the
release (as the standard library release may incorporate new features,
not just bug fixes).
User Scenarios
==============
The versioning scheme proposed above is based on a number of user scenarios
that are likely to be encountered if this scheme is adopted. In each case,
the scenario is described for both the status quo (i.e. slow release cycle)
the versioning scheme in this PEP and the free wheeling minor version number
scheme proposed in :pep:`407`.
To give away the ending, the point of using a separate version number is that
for almost all scenarios, the important number is the *language* version, not
the standard library version. Most users won't even need to care that the
standard library version number exists. In the two identified cases where
it matters, providing it as a separate number is actually clearer and more
explicit than embedding the two different kinds of number into a single
sequence and then tagging some of the numbers in the unified sequence as
special.
Novice user, downloading Python from python.org in March 2013
-------------------------------------------------------------
**Status quo:** must choose between 3.3 and 2.7
**This PEP:** must choose between 3.3 (33.1), 3.3 and 2.7.
**PEP 407:** must choose between 3.4, 3.3 (LTS) and 2.7.
**Verdict:** explaining the meaning of a Long Term Support release is about as
complicated as explaining the meaning of the proposed standard library release
version numbers. I call this a tie.
Novice user, attempting to judge currency of third party documentation
----------------------------------------------------------------------
**Status quo:** minor version differences indicate 18-24 months of
language evolution
**This PEP:** same as status quo for language core, standard library version
numbers indicate 6 months of standard library evolution.
**PEP 407:** minor version differences indicate 18-24 months of language
evolution up to 3.3, then 6 months of language evolution thereafter.
**Verdict:** Since language changes and deprecations can have a much bigger
effect on the accuracy of third party documentation than the addition of new
features to the standard library, I'm calling this a win for the scheme
in this PEP.
Novice user, looking for an extension module binary release
-----------------------------------------------------------
**Status quo:** look for the binary corresponding to the Python version you are
running.
**This PEP:** same as status quo.
**PEP 407 (full releases):** same as status quo, but corresponding binary version
is more likely to be missing (or, if it does exist, has to be found amongst
a much larger list of alternatives).
**PEP 407 (ABI updates limited to LTS releases):** all binary release pages will
need to tell users that Python 3.3, 3.4 and 3.5 all need the 3.3 binary.
**Verdict:** I call this a clear win for the scheme in this PEP. Absolutely
nothing changes from the current situation, since the standard library
version is actually irrelevant in this case (only binary extension
compatibility is important).
Extension module author, deciding whether or not to make a binary release
-------------------------------------------------------------------------
**Status quo:** unless using the :pep:`384` stable ABI, a new binary release is
needed every time the minor version number changes.
**This PEP:** same as status quo.
**PEP 407 (full releases):** same as status quo, but becomes a far more
frequent occurrence.
**PEP 407 (ABI updates limited to LTS releases):** before deciding, must first
look up whether the new release is an LTS release or an interim release. If
it is an LTS release, then a new build is necessary.
**Verdict:** I call this another clear win for the scheme in this PEP. As with
the end user facing side of this problem, the standard library version is
actually irrelevant in this case. Moving that information out to a
separate number avoids creating unnecessary confusion.
Python developer, deciding priority of eliminating a Deprecation Warning
------------------------------------------------------------------------
**Status quo:** code that triggers deprecation warnings is not guaranteed to
run on a version of Python with a higher minor version number.
**This PEP:** same as status quo
**PEP 407:** unclear, as the PEP doesn't currently spell this out. Assuming the
deprecation cycle is linked to LTS releases, then upgrading to a non-LTS
release is safe but upgrading to the next LTS release may require avoiding
the deprecated construct.
**Verdict:** another clear win for the scheme in this PEP since, once again, the
standard library version is irrelevant in this scenario.
Alternative interpreter implementor, updating with new features
---------------------------------------------------------------
**Status quo:** new Python versions arrive infrequently, but are a mish-mash of
standard library updates and core language definition and interpreter
changes.
**This PEP:** standard library updates, which are easier to integrate, are
made available more frequently in a form that is clearly and explicitly
compatible with the previous version of the language definition. This means
that, once an alternative implementation catches up to Python 3.3, they
should have a much easier time incorporating standard library features as
they happen (especially pure Python changes), leaving minor version number
updates as the only task that requires updates to their core compilation and
execution components.
**PEP 407 (full releases):** same as status quo, but becomes a far more
frequent occurrence.
**PEP 407 (language updates limited to LTS releases):** unclear, as the PEP
doesn't currently spell out a specific development strategy. Assuming a
3.3 compatibility branch is adopted (as proposed in this PEP), then the
outcome would be much the same, but the version number signalling would be
slightly less clear (since you would have to check to see if a particular
release was an LTS release or not).
**Verdict:** while not as clear cut as some previous scenarios, I'm still
calling this one in favour of the scheme in this PEP. Explicit is better than
implicit, and the scheme in this PEP makes a clear split between the two
different kinds of update rather than adding a separate "LTS" tag to an
otherwise ordinary release number. Tagging a particular version as being
special is great for communicating with version control systems and associated
automated tools, but it's a lousy way to communicate information to other
humans.
Python developer, deciding their minimum version dependency
-----------------------------------------------------------
**Status quo:** look for "version added" or "version changed" markers in the
documentation, check against ``sys.version_info``
**This PEP:** look for "version added" or "version changed" markers in the
documentation. If written as a bare Python version, such as "3.3", check
against ``sys.version_info``. If qualified with a standard library version,
such as "3.3 (33.1)", check against ``sys.stdlib_info``.
**PEP 407:** same as status quo
**Verdict:** the scheme in this PEP actually allows third party libraries to be
more explicit about their rate of adoption of standard library features. More
conservative projects will likely pin their dependency to the language
version and avoid features added in the standard library releases. Faster
moving projects could instead declare their dependency on a particular
standard library version. However, since :pep:`407` does have the advantage of
preserving the status quo, I'm calling this one for :pep:`407` (albeit with a
slim margin).
Python developers, attempting to reproduce a tracker issue
----------------------------------------------------------
**Status quo:** if not already provided, ask the reporter which version of
Python they're using. This is often done by asking for the first two lines
displayed by the interactive prompt or the value of ``sys.version``.
**This PEP:** same as the status quo (as ``sys.version`` will be updated to
also include the standard library version), but may be needed on additional
occasions (where the user knew enough to state their Python version, but that
proved to be insufficient to reproduce the fault).
**PEP 407:** same as the status quo
**Verdict:** another marginal win for :pep:`407`. The new standard library version
*is* an extra piece of information that users may need to pass back to
developers when reporting issues with Python libraries (or Python itself,
on our own tracker). However, by including it in ``sys.version``, many
fault reports will already include it, and it is easy to request if needed.
CPython release managers, handling a security fix
-------------------------------------------------
**Status quo:** create a new maintenance release incorporating the security
fix and any other bug fixes under source control. Also create source releases
for any branches open solely for security fixes.
**This PEP:** same as the status quo for maintenance branches. Also create a
new standard library release (potentially incorporating new features along
with the security fix). For security branches, create source releases for
both the former maintenance branch and the standard library update branch.
**PEP 407:** same as the status quo for maintenance and security branches,
but handling security fixes for non-LTS releases is currently an open
question.
**Verdict:** until :pep:`407` is updated to actually address this scenario, a
clear win for this PEP.
Effects
=======
Effect on development cycle
---------------------------
Similar to :pep:`407`, this PEP will break up the delivery of new features into
more discrete chunks. Instead of a whole raft of changes landing all at once
in a language release, each language release will be limited to 6 months
worth of standard library changes, as well as any changes associated with
new syntax.
Effect on workflow
------------------
This PEP proposes the creation of a single additional branch for use in the
normal workflow. After the release of 3.3, the following branches would be
in use::
2.7 # Maintenance branch, no change
3.3 # Maintenance branch, as for 3.2
3.3-compat # New branch, backwards compatible changes
default # Language changes, standard library updates that depend on them
When working on a new feature, developers will need to decide whether or not
it is an acceptable change for a standard library release. If so, then it
should be checked in on ``3.3-compat`` and then merged to ``default``.
Otherwise it should be checked in directly to ``default``.
The "version added" and "version changed" markers for any changes made on
the ``3.3-compat`` branch would need to be flagged with both the language
version and the standard library version. For example: "3.3 (33.1)".
Any changes made directly on the ``default`` branch would just be flagged
with "3.4" as usual.
The ``3.3-compat`` branch would be closed to normal development at the
same time as the ``3.3`` maintenance branch. The ``3.3-compat`` branch would
remain open for security fixes for the same period of time as the ``3.3``
maintenance branch.
Effect on bugfix cycle
----------------------
The effect on the bug fix workflow is essentially the same as that on the
workflow for new features - there is one additional branch to pass through
before the change reaches the ``default`` branch.
If critical bugs are found in a maintenance release, then new maintenance and
standard library releases will be created to resolve the problem. The final
part of the version number will be incremented for both the language version
and the standard library version.
If critical bugs are found in a standard library release that do not affect
the associated maintenance release, then only a new standard library release
will be created and only the standard library's version number will be
incremented.
Note that in these circumstances, the standard library release *may* include
additional features, rather than just containing the bug fix. It is
assumed that anyone that cares about receiving *only* bug fixes without any
new features mixed in will already be relying strictly on the maintenance
releases rather than using the new standard library releases.
Effect on the community
-----------------------
:pep:`407` has this to say about the effects on the community:
People who value stability can just synchronize on the LTS releases which,
with the proposed figures, would give a similar support cycle (both in
duration and in stability).
I believe this statement is just plain wrong. Life isn't that simple. Instead,
developers of third party modules and frameworks will come under pressure to
support the full pace of the new release cycle with binary updates, teachers
and book authors will receive complaints that they're only covering an "old"
version of Python ("You're only using 3.3, the latest is 3.5!"), etc.
As the minor version number starts climbing 3 times faster than it has in the
past, I believe perceptions of language stability would also fall (whether
such opinions were justified or not).
I believe isolating the increased pace of change to the standard library,
and clearly delineating it with a separate version number will greatly
reassure the rest of the community that no, we're not suddenly
asking them to triple their own rate of development. Instead, we're merely
going to ship standard library updates for the next language release in
6-monthly installments rather than delaying them all until the next language
definition update, even those changes that are backwards compatible with the
previously released version of Python.
The community benefits listed in :pep:`407` are equally applicable to this PEP,
at least as far as the standard library is concerned:
People who value reactivity and access to new features (without taking the
risk to install alpha versions or Mercurial snapshots) would get much more
value from the new release cycle than currently.
People who want to contribute new features or improvements would be more
motivated to do so, knowing that their contributions will be more quickly
available to normal users.
If the faster release cycle encourages more people to focus on contributing
to the standard library rather than proposing changes to the language
definition, I don't see that as a bad thing.
Handling News Updates
=====================
What's New?
-----------
The "What's New" documents would be split out into separate documents for
standard library releases and language releases. So, during the 3.3 release
cycle, we would see:
* What's New in Python 3.3?
* What's New in the Python Standard Library 33.1?
* What's New in the Python Standard Library 33.2?
* What's New in the Python Standard Library 33.3?
And then finally, we would see the next language release:
* What's New in Python 3.4?
For the benefit of users that ignore standard library releases, the 3.4
What's New would link back to the What's New documents for each of the
standard library releases in the 3.3 series.
NEWS
----
Merge conflicts on the NEWS file are already a hassle. Since this PEP
proposes introduction of an additional branch into the normal workflow,
resolving this becomes even more critical. While Mercurial phases may
help to some degree, it would be good to eliminate the problem entirely.
One suggestion from Barry Warsaw is to adopt a non-conflicting
separate-files-per-change approach, similar to that used by Twisted [2]_.
Given that the current manually updated NEWS file will be used for the 3.3.0
release, one possible layout for such an approach might look like::
Misc/
NEWS # Now autogenerated from news_entries
news_entries/
3.3/
NEWS # Original 3.3 NEWS file
maint.1/ # Maintenance branch changes
core/
<news entries>
builtins/
<news entries>
extensions/
<news entries>
library/
<news entries>
documentation/
<news entries>
tests/
<news entries>
compat.1/ # Compatibility branch changes
builtins/
<news entries>
extensions/
<news entries>
library/
<news entries>
documentation/
<news entries>
tests/
<news entries>
# Add maint.2, compat.2 etc as releases are made
3.4/
core/
<news entries>
builtins/
<news entries>
extensions/
<news entries>
library/
<news entries>
documentation/
<news entries>
tests/
<news entries>
# Add maint.1, compat.1 etc as releases are made
Putting the version information in the directory hierarchy isn't strictly
necessary (since the NEWS file generator could figure out from the version
history), but does make it easier for *humans* to keep the different versions
in order.
Other benefits of reduced version coupling
==========================================
Slowing down the language release cycle
---------------------------------------
The current release cycle is a compromise between the desire for stability
in the core language definition and C extension ABI, and the desire to get
new features (most notably standard library updates) into user's hands more
quickly.
With the standard library release cycle decoupled (to some degree) from that
of the core language definition, it provides an opportunity to actually
*slow down* the rate of change in the language definition. The language
moratorium for Python 3.2 effectively slowed that cycle down to *more than 3
years* (3.1: June 2009, 3.3: August 2012) without causing any major
problems or complaints.
The NEWS file management scheme described above is actually designed to
allow us the flexibility to slow down language releases at the same time
as standard library releases become more frequent.
As a simple example, if a full two years was allowed between 3.3 and 3.4,
the 3.3 release cycle would end up looking like::
3.2.4 # Maintenance release
3.3.0 # Language release
3.3.1 # Maintenance release
3.3 (33.1) # Standard library release
3.3.2 # Maintenance release
3.3 (33.2) # Standard library release
3.3.3 # Maintenance release
3.3 (33.3) # Standard library release
3.3.4 # Maintenance release
3.3 (33.4) # Standard library release
3.4.0 # Language release
The elegance of the proposed branch structure and NEWS entry layout is that
this decision wouldn't really need to be made until shortly before the planned
3.4 release date. At that point, the decision could be made to postpone the
3.4 release and keep the ``3.3`` and ``3.3-compat`` branches open after the
3.3.3 maintenance release and the 3.3 (33.3) standard library release, thus
adding another standard library release to the cycle. The choice between
another standard library release or a full language release would then be
available every 6 months after that.
Further increasing the pace of standard library development
-----------------------------------------------------------
As noted in the previous section, one benefit of the scheme proposed in this
PEP is that it largely decouples the language release cycle from the
standard library release cycle. The standard library could be updated every
3 months, or even once a month, without having any flow on effects on the
language version numbering or the perceived stability of the core language.
While that pace of development isn't practical as long as the binary
installer creation for Windows and Mac OS X involves several manual steps
(including manual testing) and for as long as we don't have separate
"<branch>-release" trees that only receive versions that have been marked as
good by the stable buildbots, it's still a useful criterion to keep in mind
when considering proposed new versioning schemes: what if we eventually want
to make standard library releases even *faster* than every 6 months?
If the practical issues were ever resolved, then the separate standard
library versioning scheme in this PEP could handle it. The tagged version
number approach proposed in :pep:`407` could not (at least, not without a lot
of user confusion and uncertainty).
Other Questions
===============
Why not use the major version number?
-------------------------------------
The simplest and most logical solution would actually be to map the
major.minor.micro version numbers to the language version, stdlib version
and maintenance release version respectively.
Instead of releasing Python 3.3.0, we would instead release Python 4.0.0
and the release cycle would look like::
4.0.0 # Language release
4.0.1 # Maintenance release
4.1.0 # Standard library release
4.0.2 # Maintenance release
4.2.0 # Standard library release
4.0.3 # Maintenance release
4.3.0 # Standard library release
5.0.0 # Language release
However, the ongoing pain of the Python 2 -> Python 3 transition (and
associated workarounds like the ``python3`` and ``python2`` symlinks to
refer directly to the desired release series) means that this simple option
isn't viable for historical reasons.
One way that this simple approach *could* be made to work is to merge the
current major and minor version numbers directly into a 2-digit major
version number::
33.0.0 # Language release
33.0.1 # Maintenance release
33.1.0 # Standard library release
33.0.2 # Maintenance release
33.2.0 # Standard library release
33.0.3 # Maintenance release
33.3.0 # Standard library release
34.0.0 # Language release
Why not use a four part version number?
---------------------------------------
Another simple versioning scheme would just add a "standard library" version
into the existing versioning scheme::
3.3.0.0 # Language release
3.3.0.1 # Maintenance release
3.3.1.0 # Standard library release
3.3.0.2 # Maintenance release
3.3.2.0 # Standard library release
3.3.0.3 # Maintenance release
3.3.3.0 # Standard library release
3.4.0.0 # Language release
However, this scheme isn't viable due to backwards compatibility constraints
on the ``sys.version_info`` structure.
Why not use a date-based versioning scheme?
-------------------------------------------
Earlier versions of this PEP proposed a date-based versioning scheme for
the standard library. However, such a scheme made it very difficult to
handle out-of-cycle releases to fix security issues and other critical
bugs in standard library releases, as it required the following steps:
1. Change the release version number to the date of the current month.
2. Update the What's New, NEWS and documentation to refer to the new release
number.
3. Make the new release.
With the sequential scheme now proposed, such releases should at most require
a little tidying up of the What's New document before making the release.
Why isn't PEP 384 enough?
-------------------------
:pep:`384` introduced the notion of a "Stable ABI" for CPython, a limited
subset of the full C ABI that is guaranteed to remain stable. Extensions
built against the stable ABI should be able to support all subsequent
Python versions with the same binary.
This will help new projects to avoid coupling their C extension modules too
closely to a specific version of CPython. For existing modules, however,
migrating to the stable ABI can involve quite a lot of work (especially for
extension modules that define a lot of classes). With limited development
resources available, any time spent on such a change is time that could
otherwise have been spent working on features that offer more direct benefits
to end users.
There are also other benefits to separate versioning (as described above)
that are not directly related to the question of binary compatibility with
third party C extensions.
Why no binary compatible additions to the C ABI in standard library releases?
-----------------------------------------------------------------------------
There's a case to be made that *additions* to the CPython C ABI could
reasonably be permitted in standard library releases. This would give C
extension authors the same freedom as any other package or module author
to depend either on a particular language version or on a standard library
version.
The PEP currently associates the interpreter version with the language
version, and therefore limits major interpreter changes (including C ABI
additions) to the language releases.
An alternative, internally consistent, approach would be to link the
interpreter version with the standard library version, with only changes that
may affect backwards compatibility limited to language releases.
Under such a scheme, the following changes would be acceptable in standard
library releases:
* Standard library updates
* new features in pure Python modules
* new features in C extension modules (subject to :pep:`399` compatibility
requirements)
* new features in language builtins
* Interpreter implementation updates
* binary compatible additions to the C ABI
* changes to the compilation toolchain that do not affect the AST or alter
the bytecode magic number
* changes to the core interpreter eval loop
* bug fixes from the corresponding maintenance release
And the following changes would be acceptable in language releases:
* new language syntax
* any updates acceptable in a standard library release
* new deprecation warnings
* removal of previously deprecated features
* changes to the AST
* changes to the emitted bytecode that require altering the magic number
* binary incompatible changes to the C ABI (although the :pep:`384` stable ABI
must still be preserved)
While such an approach could probably be made to work, there does not appear
to be a compelling justification for it, and the approach currently described
in the PEP is simpler and easier to explain.
Why not separate out the standard library entirely?
---------------------------------------------------
A concept that is occasionally discussed is the idea of making the standard
library truly independent from the CPython reference implementation.
My personal opinion is that actually making such a change would involve a
lot of work for next to no pay-off. CPython without the standard library is
useless (the build chain won't even run, let alone the test suite). You also
can't create a standalone pure Python standard library either, because too
many "standard library modules" are actually tightly linked in to the
internal details of their respective interpreters (for example, the builtins,
``weakref``, ``gc``, ``sys``, ``inspect``, ``ast``).
Creating a separate CPython development branch that is kept compatible with
the previous language release, and making releases from that branch that are
identified with a separate standard library version number should provide
most of the benefits of a separate standard library repository with only a
fraction of the pain.
Acknowledgements
================
Thanks go to the :pep:`407` authors for starting this discussion, as well as
to those authors and Larry Hastings for initial discussions of the proposal
made in this PEP.
References
==========
.. [2] Twisted's "topfiles" approach to NEWS generation
https://web.archive.org/web/20120305142914/http://twistedmatrix.com/trac/wiki/ReviewProcess#Newsfiles
Copyright
=========
This document has been placed in the public domain.