|Author:||larry at hastings.org (Larry Hastings)|
|Discussions-To:||Python-Dev <python-dev at python.org>|
Python's descriptor protocol requires that descriptors be members of the type of an object. This PEP proposes an extension to the descriptor protocol allowing use of the descriptor protocol for members of instances. This would permit using properties in modules.
Python's descriptor protocol guides programmers towards elegant API design. If your class supports a data-like member, and you might someday need to run code when changing the member's value, you're encouraged to simply declare it as a simple data member of the class for now. If in the future you do need to run code, you can change it to a "property", and happily the API doesn't change.
But consider this second bit of best-practice Python API design: if you're writing a singleton, don't write a class, just build your code directly into a module. Don't make your users instantiate a singleton class, don't make your users have to dereference through a singleton object stored in a module, just have module-level functions and module-level data.
Unfortunately these two best practices are in opposition. The problem is that properties aren't supported on modules. Modules are instances of a single generic module type, and it's not feasible to modify or subclass this type to add a property to one's module. This means that programmers facing this API design decision, where the data-like member is a singleton stored in a module, must preemptively add ugly "getters" and "setters" for the data.
Adding support for module properties in pure Python has recently become possible; as of Python 3.5, Python permits assigning to the __class__ attribute of module objects, specifically for this purpose. Here's an example of using this functionality to add a property to a module:
import sys, types class _MyModuleType(types.ModuleType): @property def prop(self, instance, owner): ... sys.modules[__name__].__class__ = _MyModuleType
This works, and is supported behavior, but it's clumsy and obscure.
This PEP proposes a per-type opt-in extension to the descriptor protocol specifically designed to enable properties in modules. The mechanism is a way to honor the descriptor protocol for members of instances of a class without the member being declared as a class variable.
Although this is being proposed as a general mechanism, the author currently only foresees this as being useful for module objects.
The basic idea is simple: modify the tp_descr_get and tp_descr_set functions exposed by PyModule_Type to inspect the attribute interacted with, and if it supports the descriptor protocol, call the relevant exposed function.
Our implementation faces two challenges:
- Since this code will be run every time an attribute is looked up on a method, it needs to add very little overhead in the general case, where the object stored in the attribute is not a descriptor.
- Since functions are descriptors, we must take care to not honor the descriptor protocol for all objects. Otherwise, all module-level functions will suddenly become bound to the module instance as if they were method calls on the module object. The module handle would be passed in as a "self" argument to all module-level functions.
Both challenges can be solved with the same approach: we define a new "fast subclass" flag that means "This object is a descriptor, and it should be honored directly when this object is looked up as an attribute of an instance". So far this flag is only set on two types: property and collections.abc.InstanceDescriptor. The latter is an abstract base class, whose only purpose is to allow user classes to inherit this "fast subclass" flag.
Armin Rigo essentially proposed this mechanism when presented with the idea of "module properties", and educated the author both on the complexities of the problem and the proper solution. Nathaniel J. Smith pointed out the 3.5 extension about assigning to __class__ on module objects, and provided the example.
This document has been placed in the public domain.