|Title:||Preserving Class Attribute Definition Order|
|Author:||Eric Snow <ericsnowcurrently at gmail.com>|
|Post-History:||7-Jun-2016, 11-Jun-2016, 20-Jun-2016, 24-Jun-2016|
- Why a tuple?
- Why not a read-only attribute?
- Why not "__attribute_order__"?
- Why not ignore "dunder" names?
- Why None instead of an empty tuple?
- Why None instead of not setting the attribute?
- Why constrain manually set values?
- Why not hide __definition_order__ on non-type objects?
- What about __slots__?
- Why is __definition_order__ even necessary?
- Support for C-API Types
- Other Python Implementations
Since compact dict has landed in 3.6, __definition_order__ has been removed. cls.__dict__ now mostly accomplishes the same thing instead.
The class definition syntax is ordered by its very nature. Class attributes defined there are thus ordered. Aside from helping with readability, that ordering is sometimes significant. If it were automatically available outside the class definition then the attribute order could be used without the need for extra boilerplate (such as metaclasses or manually enumerating the attribute order). Given that this information already exists, access to the definition order of attributes is a reasonable expectation. However, currently Python does not preserve the attribute order from the class definition.
This PEP changes that by preserving the order in which attributes are introduced in the class definition body. That order will now be preserved in the __definition_order__ attribute of the class. This allows introspection of the original definition order, e.g. by class decorators.
Additionally, this PEP requires that the default class definition namespace be ordered (e.g. OrderedDict) by default. The long- lived class namespace (__dict__) will remain a dict.
The attribute order from a class definition may be useful to tools that rely on name order. However, without the automatic availability of the definition order, those tools must impose extra requirements on users. For example, use of such a tool may require that your class use a particular metaclass. Such requirements are often enough to discourage use of the tool.
Some tools that could make use of this PEP include:
- documentation generators
- testing frameworks
- CLI frameworks
- web frameworks
- config generators
- data serializers
- enum factories (my original motivation)
When a class is defined using a class statement, the class body is executed within a namespace. Currently that namespace defaults to dict. If the metaclass defines __prepare__() then the result of calling it is used for the class definition namespace.
After the execution completes, the definition namespace is copied into a new dict. Then the original definition namespace is discarded. The new copy is stored away as the class's namespace and is exposed as __dict__ through a read-only proxy.
The class attribute definition order is represented by the insertion order of names in the definition namespace. Thus, we can have access to the definition order by switching the definition namespace to an ordered mapping, such as collections.OrderedDict. This is feasible using a metaclass and __prepare__, as described above. In fact, exactly this is by far the most common use case for using __prepare__.
At that point, the only missing thing for later access to the definition order is storing it on the class before the definition namespace is thrown away. Again, this may be done using a metaclass. However, this means that the definition order is preserved only for classes that use such a metaclass. There are two practical problems with that:
First, it requires the use of a metaclass. Metaclasses introduce an extra level of complexity to code and in some cases (e.g. conflicts) are a problem. So reducing the need for them is worth doing when the opportunity presents itself. PEP 422 and PEP 487 discuss this at length. We have such an opportunity by using an ordered mapping (e.g. OrderedDict for CPython at least) for the default class definition namespace, virtually eliminating the need for __prepare__().
Second, only classes that opt in to using the OrderedDict-based metaclass will have access to the definition order. This is problematic for cases where universal access to the definition order is important.
- all classes have a __definition_order__ attribute
- __definition_order__ is a tuple of identifiers (or None)
- __definition_order__ is always set:
- during execution of the class body, the insertion order of names into the class definition namespace is stored in a tuple
- if __definition_order__ is defined in the class body then it must be a tuple of identifiers or None; any other value will result in TypeError
- classes that do not have a class definition (e.g. builtins) have their __definition_order__ set to None
- classes for which __prepare__() returned something other than OrderedDict (or a subclass) have their __definition_order__ set to None (except where #2 applies)
- dir() will not depend on __definition_order__
- descriptors and custom __getattribute__ methods are unconstrained regarding __definition_order__
- the default class definition namespace is now an ordered mapping (e.g. OrderdDict)
- cls.__dict__ does not change, remaining a read-only proxy around dict
Note that Python implementations which have an ordered dict won't need to change anything.
The following code demonstrates roughly equivalent semantics for both parts 1 and 2:
class Meta(type): @classmethod def __prepare__(cls, *args, **kwargs): return OrderedDict() class Spam(metaclass=Meta): ham = None eggs = 5 __definition_order__ = tuple(locals())
Use of a tuple reflects the fact that we are exposing the order in which attributes on the class were defined. Since the definition is already complete by the time __definition_order__ is set, the content and order of the value won't be changing. Thus we use a type that communicates that state of immutability.
There are some valid arguments for making __definition_order__ a read-only attribute (like cls.__dict__ is). Most notably, a read-only attribute conveys the nature of the attribute as "complete", which is exactly correct for __definition_order__. Since it represents the state of a particular one-time event (execution of the class definition body), allowing the value to be replaced would reduce confidence that the attribute corresponds to the original class body. Furthermore, often an immutable-by-default approach helps to make data easier to reason about.
However, in this case there still isn't a strong reason to counter the well-worn precedent found in Python. Per Guido:
I don't see why it needs to be a read-only attribute. There are very few of those -- in general we let users play around with things unless we have a hard reason to restrict assignment (e.g. the interpreter's internal state could be compromised). I don't see such a hard reason here.
Also, note that a writeable __definition_order__ allows dynamically created classes (e.g. by Cython) to still have __definition_order__ properly set. That could certainly be handled through specific class- creation tools, such as type() or the C-API, without the need to lose the semantics of a read-only attribute. However, with a writeable attribute it's a moot point.
__definition_order__ is centered on the class definition body. The use cases for dealing with the class namespace (__dict__) post-definition are a separate matter. __definition_order__ would be a significantly misleading name for a feature focused on more than class definition.
Names starting and ending with "__" are reserved for use by the interpreter. In practice they should not be relevant to the users of __definition_order__. Instead, for nearly everyone they would only be clutter, causing the same extra work (filtering out the dunder names) for the majority. In cases where a dunder name is significant, the class definition could manually set __definition_order__, making the common case simpler.
However, leaving dunder names out of __definition_order__ means that their place in the definition order would be unrecoverably lost. Dropping dunder names by default may inadvertently cause problems for classes that use dunder names unconventionally. In this case it's better to play it safe and preserve all the names from the class definition. This isn't a big problem since it is easy to filter out dunder names:
(name for name in cls.__definition_order__ if not (name.startswith('__') and name.endswith('__')))
In fact, in some application contexts there may be other criteria on which similar filtering would be applied, such as ignoring any name starting with "_", leaving out all methods, or including only descriptors. Ultimately dunder names aren't a special enough case to be treated exceptionally.
Note that a couple of dunder names (__name__ and __qualname__) are injected by default by the compiler. So they will be included even though they are not strictly part of the class definition body.
A key objective of adding __definition_order__ is to preserve information in class definitions which was lost prior to this PEP. One consequence is that __definition_order__ implies an original class definition. Using None allows us to clearly distinguish classes that do not have a definition order. An empty tuple clearly indicates a class that came from a definition statement but did not define any attributes there.
The absence of an attribute requires more complex handling than None does for consumers of __definition_order__.
If __definition_order__ is manually set in the class body then it will be used. We require it to be a tuple of identifiers (or None) so that consumers of __definition_order__ may have a consistent expectation for the value. That helps maximize the feature's usefulness.
We could also also allow an arbitrary iterable for a manually set __definition_order__ and convert it into a tuple. However, not all iterables infer a definition order (e.g. set). So we opt in favor of requiring a tuple.
Python doesn't make much effort to hide class-specific attributes during lookup on instances of classes. While it may make sense to consider __definition_order__ a class-only attribute, hidden during lookup on objects, setting precedent in that regard is beyond the goals of this PEP.
__slots__ will be added to __definition_order__ like any other name in the class definition body. The actual slot names will not be added to __definition_order__ since they aren't set as names in the definition namespace.
Since the definition order is not preserved in __dict__, it is lost once class definition execution completes. Classes could explicitly set the attribute as the last thing in the body. However, then independent decorators could only make use of classes that had done so. Instead, __definition_order__ preserves this one bit of info from the class body so that it is universally available.
Arguably, most C-defined Python types (e.g. built-in, extension modules) have a roughly equivalent concept of a definition order. So conceivably __definition_order__ could be set for such types automatically. This PEP does not introduce any such support. However, it does not prohibit it either. However, since __definition_order__ can be set at any time through normal attribute assignment, it does not need any special treatment in the C-API.
The specific cases:
- builtin types
This PEP does not break backward compatibility, except in the case that someone relies strictly on dict as the class definition namespace. This shouldn't be a problem since issubclass(OrderedDict, dict) is true.
In addition to the class syntax, the following expose the new behavior:
Also, the 3-argument form of builtins.type() will allow inclusion of __definition_order__ in the namespace that gets passed in. It will be subject to the same constraints as when __definition_order__ is explicitly defined in the class body.
Pending feedback, the impact on Python implementations is expected to be minimal. All conforming implementations are expected to set __definition_order__ as described in this PEP.
Instead of storing the definition order in __definition_order__, the now-ordered definition namespace could be copied into a new OrderedDict. This would then be used as the mapping proxied as __dict__. Doing so would mostly provide the same semantics.
However, using OrderedDict for __dict__ would obscure the relationship with the definition namespace, making it less useful.
Additionally, (in the case of OrderedDict specifically) doing this would require significant changes to the semantics of the concrete dict C-API.
There has been some discussion about moving to a compact dict implementation which would (mostly) preserve insertion order. However the lack of an explicit __definition_order__ would still remain as a pain point.
PEP 422 introduced a new "namespace" keyword arg to class definitions that effectively replaces the need to __prepare__(). [pep422] However, the proposal was withdrawn in favor of the simpler PEP 487.
This has all the same problems as writing your own metaclass. The only advantage is that you don't have to actually write this metaclass. So it doesn't offer any benefit in the context of this PEP.
Each class's __qualname__ is determined at compile-time. This same concept could be applied to __definition_order__. The result of composing __definition_order__ at compile-time would be nearly the same as doing so at run-time.
Comparative implementation difficulty aside, the key difference would be that at compile-time it would not be practical to preserve definition order for attributes that are set dynamically in the class body (e.g. locals()[name] = value). However, they should still be reflected in the definition order. One possible resolution would be to require class authors to manually set __definition_order__ if they define any class attributes dynamically.
Ultimately, the use of OrderedDict at run-time or compile-time discovery is almost entirely an implementation detail.
|[impl]||issue #24254 (https://bugs.python.org/issue24254)|
|[nick_concern]||Nick's concerns about mutability (https://mail.python.org/pipermail/python-dev/2016-June/144883.html)|
|[pep422]||PEP 422 (https://www.python.org/dev/peps/pep-0422/#order-preserving-classes)|
|[pep487]||PEP 487 (https://www.python.org/dev/peps/pep-0487/#defining-arbitrary-namespaces)|
|[orig]||original discussion (https://mail.python.org/pipermail/python-ideas/2013-February/019690.html)|
|[followup1]||follow-up 1 (https://mail.python.org/pipermail/python-dev/2013-June/127103.html)|
|[followup2]||follow-up 2 (https://mail.python.org/pipermail/python-dev/2015-May/140137.html)|
This document has been placed in the public domain.