|Title:||A standard mechanism for backward compatibility|
|Author:||Ed Schofield <ed at pythoncharmers.com>|
|BDFL-Delegate:||Brett Cannon (on behalf of the steering council)|
The steering council decided that the __past__ aspect of this proposal was too complicated for the potential benefit. The other aspect of stronger requirements for backwards-compatibility should be addressed by PEP 387.
This PEP is complementary to PEPs 5, 236, and 387, and shares similar goals.
This PEP explains the need for an additional compatibility mechanism in support of PEP 5, "Guidelines for Language Evolution". PEP 236, "Back to the __future__", introduced a mechanism for forward compatibility in support of PEP 5 but noted that a new mechanism for backward compatibility was outside the scope of that PEP. A related PEP (in progress) introduces such a mechanism for backward compatibility.
From PEP 236: "From time to time, Python makes an incompatible change to the advertised semantics of core language constructs, or changes their accidental (implementation-dependent) behavior in some way. While this is never done capriciously, and is always done with the aim of improving the language over the long term, over the short term it's contentious and disrupting. PEP 5, Guidelines for Language Evolution, suggests ways to ease the pain, and this PEP  introduces some machinery in support of that."
Also from PEP 236: "The purpose of future_statement is to make life easier for people who keep current with the latest release in a timely fashion. We don't hate you if you don't, but your problems are much harder to solve, and somebody with those problems will need to write a PEP addressing them. future_statement is aimed at a different audience."
When an incompatible change to core language syntax or semantics is being made, Python currently provides the future_statement mechanism for providing forward compatibility until the release that enforces the new syntax or semantics, but provides no corresponding standard mechanism for providing backward compatibility after this release.
A consequence of this asymmetry is that, with respect to a breaking change, the older (pre-breaking) version of the Python interpreter is more capable than the newer (breaking) version; the older interpreter can use both code designed prior to the change and newer code, whereas the newer interpreter is only capable of using code that has been upgraded to support the changed feature.
As an example, consider the changes to the division operator introduced in PEP 238 in 2001, soon after PEP 236 introduced the future_statement mechanism. PEP 238 outlines a suite of useful forward-compatibility mechanisms for "true division" in the Python 2.x series but omits to include any backward-compatibility mechanisms for after "true division" was first enforced in Python 3.0. Python versions since 3.0 do not provide a backward compatibility mechanism such as from __past__ import division for code that expects the old "classic division" semantics, whereas Python versions prior to 3.0 do support both "classic division" code and also forward compatibility with code expecting "true division". A further consequence of this is that the "most compatible" interpreter with respect to the variety of division-related Python code in the wild is Python 2.7, the version before the breaking change was first enforced.
In contrast to this situation, newer versions of application software such as office suites tend to be more capable than earlier versions with respect to their support for loading different versions of their data file formats. The pattern is usually that the newer application versions can transparently load data from either their newer or their older data formats, and that the newer version defaults to saving data in the newer format. Newer application software versions tend to be backward-compatible by default. Forward compatibility is relatively rare.
This policy puts the user of the newer application software at an advantage over the user of the older software, which is usually incapable of loading data in the newer format. Sometimes it is possible for a user of a newer software application version to export data in an older version by choosing this option explicitly. In these cases, the forward-compatibility this enables may or may not be perfect; some features may be missing or the results may be otherwise suboptimal. Upgrading is therefore easy, whereas downgrading is harder.
The emergent behaviour over many users from such a policy of new attractive features plus backward compatibility features is that a natural pressure builds up on each individual user to upgrade his or her own application version, and, the more other users an individual exchanges data files with, the more acute this pressure becomes.
This PEP makes two specific, related proposals. The first is that:
PEP 5 be augmented with a 6th step in the section "Steps for Introducing Backwards-Incompatible Features" to indicate that, 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. Furthermore, PEP 387, "Backwards Compatibility Policy" (if accepted) would be augmented with the same 6th step.
As an example of how this PEP is to be applied, if the latest revision of the "true division" PEP (238) were proposed today, it would be considered incomplete. PEP 238 notes the "severe backwards compatibility issues" raised by the proposal and describes several measures for forward compatibility in the Abstract and API Changes sections. It also mentions some backward compatibility ideas raised on c.l.py, including "Use from __past__ import division to use classic division semantics in a module", but it does not put forward any backward compatibility plan as part of the proposal.
If this PEP is accepted, it would be expected that a proposal such as PEP 238, because of its large-scale compatibility implications, would also be accompanied by a backward compatibility plan that enables users of future Python versions after the breaking change has come into effect to re-enable the classic division behaviour easily in their code.
The second proposal is that:
Python provide a standard backward compatibility mechanism in parallel to the __future__ module mechanism for forward compatibility.
For reference, this document will refer to this as a "__past__" mechanism hereon, although it need not have all the characteristics of the __future__ module and future_statement mechanism.
The specific form and implementation of the __past__ mechanism is the subject of a separate PEP (in progress). However, this PEP recommends that this __past__ mechanism be designed to meet similar criteria to those outlined in PEP 296 for __future__. Specifically:
a. It should enable individual modules to specify obsolete behaviours to re-enable from older Python versions on a module-by-module basis.
b. It should be flexible enough for both Python 3.6+ and point releases of earlier versions to reintroduce backward compatibility with older Python syntax or semantics for user modules that invoke the __past__ mechanism.
c. It should be possible to run older code augmented to invoke __past__ behaviours on older Python versions such as 2.x that have no knowledge of the specific __past__ features invoked, or even that the __past__ mechanism for backward-compatibility exists.
Some implementations of __past__ mechanisms that would violate these criteria are:
a. Import hooks. These would normally fail to work on a module-by-module basis; instead they apply recursively to all new modules imported from within a module.
b. A new piece of syntax or new semantics for Python 3.6 that is incompatible with prior versions.
c. A function added in Python 3.6 to a module in the Python standard library that exists under the same name in prior Python versions.
The benefit to Python-dev of adopting this proposal is that future backward-incompatible changes can be less disruptive if these changes each have a corresponding __past__ feature that has been implemented and can be invoked easily by users of future Python versions. This can help the language to evolve more quickly and more effectively to correct for design mistakes.
The benefit to conservative users is obvious: they can add support for the latest shiny compatibility-breaking Python version to their code merely by adding a __past__ incantation (perhaps a single line) to each module, and that this can be automated. They can then upgrade their interpreter to the latest version and gain access to the latest shiny Python features.
The benefit to the community is that, if ten thousand users rely on package XYZ, and package XYZ can trivially add support for the latest Python version, those ten thousand users can also upgrade to the latest Python version quickly, without being held back waiting for package XYZ to do this.
Q1: Does this PEP require that Python keep two possible sets of semantics for each backward-incompatible feature forever?
A1: Definitely not. Legacy features can still be phased out when appropriate -- that is, when the majority of the user-base has migrated to the newer Python version. This PEP merely proposes to shift the emphasis of the development effort directed at compatibility from 100% forwards to at least 50% backwards. Backwards compatibility is the more powerful of the two concepts for allowing a user-base to adopt the latest Python interpreter version.
Notice that it has been a long time since most users have cared about backwards compatibility for non-nested scopes, because most users have moved comfortably past Python 2.1.
Q2: But Python-dev is already overwhelmed and doesn't have the bandwidth to implement / maintain the additional complexity!
A2: Python-dev can ask the community of developers to step up and maintain backward compatibility in Python for legacy language features they care about. When the community stops caring about a particular obsolete behaviour, Python-dev can stop caring too.
The __past__ mechanism could possibly be designed to be extensible by the community, e.g. as a standard but "blessed" PyPI package, to reduce the load on the core developers.
Q3: Won't backward compatibility features lead to lots of cruft and bloat and baggage in Python?
A3: Not necessarily. First, proposals for new compatibility-breaking features in Python could be evaluated partly on the simplicity and maintainability of the implementation of their associated __past__ feature up-front.
Second, some old features are simple to provide backward compatibility for. Consider the "classic division" behaviour before Python 3.0. The python-future project contains a compatible implementation of classic division in the function future.utils.old_div:
def old_div(a, b): """ Equivalent to ``a / b`` on Python 2 without ``from __future__ import division``. """ if isinstance(a, numbers.Integral) and isinstance(b, numbers.Integral): return a // b else: return a / b
Bundling such a function with Python 3.x versions, together with a simple mechanism to invoke it for every appearance of a / b after an appropriate __past__ invocation, need not be onerous.
Q4: What about performance? Won't the performance of newer Python versions suffer under the weight of legacy features?
A4: This can be evaluated on a case-by-case basis. The major potential concern is that the performance with the new default behaviour does not suffer unduly because of the presence of the legacy option. The performance under the influence of the __past__ invocation is of secondary importance.
This document has been placed in the public domain.