|Last-Modified:||2010-01-21 01:23:17 +0000 (Thu, 21 Jan 2010)|
|Author:||Collin Winter <collinwinter at google.com>|
- Rejection Notice
- Performing Your Role
- Defining New Roles
- Relationship to Abstract Base Classes
- Open Issues
Python's existing object model organizes objects according to their implementation. It is often desirable -- especially in duck typing-based language like Python -- to organize objects by the part they play in a larger system (their intent), rather than by how they fulfill that part (their implementation). This PEP introduces the concept of roles, a mechanism for organizing objects according to their intent rather than their implementation.
In the beginning were objects. They allowed programmers to marry function and state, and to increase code reusability through concepts like polymorphism and inheritance, and lo, it was good. There came a time, however, when inheritance and polymorphism weren't enough. With the invention of both dogs and trees, we were no longer able to be content with knowing merely, "Does it understand 'bark'?" We now needed to know what a given object thought that "bark" meant.
One solution, the one detailed here, is that of roles, a mechanism orthogonal and complementary to the traditional class/instance system. Whereas classes concern themselves with state and implementation, the roles mechanism deals exclusively with the behaviours embodied in a given class.
This system was originally called "traits" and implemented for Squeak Smalltalk . It has since been adapted for use in Perl 6  where it is called "roles", and it is primarily from there that the concept is now being interpreted for Python 3. Python 3 will preserve the name "roles".
In a nutshell: roles tell you what an object does, classes tell you how an object does it.
In this PEP, I will outline a system for Python 3 that will make it possible to easily determine whether a given object's understanding of "bark" is tree-like or dog-like. (There might also be more serious examples.)
A syntax proposals in this PEP are tentative and should be considered to be strawmen. The necessary bits that this PEP depends on -- namely PEP 3115's class definition syntax and PEP 3129's class decorators -- are still being formalized and may change. Function names will, of course, be subject to lengthy bikeshedding debates.
Let's start out by defining Tree and Dog classes
class Tree(Vegetable): def bark(self): return self.is_rough() class Dog(Animal): def bark(self): return self.goes_ruff()
While both implement a bark() method with the same signature, they do wildly different things. We need some way of differentiating what we're expecting. Relying on inheritance and a simple isinstance() test will limit code reuse and/or force any dog-like classes to inherit from Dog, whether or not that makes sense. Let's see if roles can help.
@perform_role(Doglike) class Dog(Animal): ... @perform_role(Treelike) class Tree(Vegetable): ... @perform_role(SitThere) class Rock(Mineral): ...
We use class decorators from PEP 3129 to associate a particular role or roles with a class. Client code can now verify that an incoming object performs the Doglike role, allowing it to handle Wolf, LaughingHyena and Aibo  instances, too.
Roles can be composed via normal inheritance:
@perform_role(Guard, MummysLittleDarling) class GermanShepherd(Dog): def guard(self, the_precious): while True: if intruder_near(the_precious): self.growl() def get_petted(self): self.swallow_pride()
Here, GermanShepherd instances perform three roles: Guard and MummysLittleDarling are applied directly, whereas Doglike is inherited from Dog.
Roles can be assigned at runtime, too, by unpacking the syntactic sugar provided by decorators.
Say we import a Robot class from another module, and since we know that Robot already implements our Guard interface, we'd like it to play nicely with guard-related code, too.
This takes effect immediately and impacts all instances of Robot.
Just because we've told our robot army that they're guards, we'd like to check in on them occasionally and make sure they're still at their task.
>>> performs(our_robot, Guard) True
What about that one robot over there?
>>> performs(that_robot_over_there, Guard) True
The performs() function is used to ask if a given object fulfills a given role. It cannot be used, however, to ask a class if its instances fulfill a role:
>>> performs(Robot, Guard) False
This is because the Robot class is not interchangeable with a Robot instance.
Roles are defined like a normal class, but use the Role metaclass.
class Doglike(metaclass=Role): ...
Metaclasses are used to indicate that Doglike is a Role in the same way 5 is an int and tuple is a type.
Roles may inherit from other roles; this has the effect of composing them. Here, instances of Dog will perform both the Doglike and FourLegs roles.
class FourLegs(metaclass=Role): pass class Doglike(FourLegs, Carnivor): pass @perform_role(Doglike) class Dog(Mammal): pass
So far we've only defined empty roles -- not very useful things. Let's now require that all classes that claim to fulfill the Doglike role define a bark() method:
class Doglike(FourLegs): def bark(self): pass
No decorators are required to flag the method as "abstract", and the method will never be called, meaning whatever code it contains (if any) is irrelevant. Roles provide only abstract methods; concrete default implementations are left to other, better-suited mechanisms like mixins.
Once you have defined a role, and a class has claimed to perform that role, it is essential that that claim be verified. Here, the programmer has misspelled one of the methods required by the role.
@perform_role(FourLegs) class Horse(Mammal): def run_like_teh_wind(self) ...
This will cause the role system to raise an exception, complaining that you're missing a run_like_the_wind() method. The role system carries out these checks as soon as a class is flagged as performing a given role.
Concrete methods are required to match exactly the signature demanded by the role. Here, we've attempted to fulfill our role by defining a concrete version of bark(), but we've missed the mark a bit.
@perform_role(Doglike) class Coyote(Mammal): def bark(self, target=moon): pass
This method's signature doesn't match exactly with what the Doglike role was expecting, so the role system will throw a bit of a tantrum.
The following are strawman proposals for how roles might be expressed in Python. The examples here are phrased in a way that the roles mechanism may be implemented without changing the Python interpreter. (Examples adapted from an article on Perl 6 roles by Curtis Poe .)
Static class role assignment
@perform_role(Thieving) class Elf(Character): ...
perform_role() accepts multiple arguments, such that this is also legal:
@perform_role(Thieving, Spying, Archer) class Elf(Character): ...
The Elf class now performs both the Thieving, Spying, and Archer roles.
if performs(my_elf, Thieving): ...
The second argument to performs() may also be anything with a __contains__() method, meaning the following is legal:
if performs(my_elf, set([Thieving, Spying, BoyScout])): ...
Like isinstance(), the object needs only to perform a single role out of the set in order for the expression to be true.
Early drafts of this PEP  envisioned roles as competing with the abstract base classes proposed in PEP 3119. After further discussion and deliberation, a compromise and a delegation of responsibilities and use-cases has been worked out as follows:
Roles provide a way of indicating a object's semantics and abstract capabilities. A role may define abstract methods, but only as a way of delineating an interface through which a particular set of semantics are accessed. An Ordering role might require that some set of ordering operators be defined.
class Ordering(metaclass=Role): def __ge__(self, other): pass def __le__(self, other): pass def __ne__(self, other): pass # ...and so on
In this way, we're able to indicate an object's role or function within a larger system without constraining or concerning ourselves with a particular implementation.
Abstract base classes, by contrast, are a way of reusing common, discrete units of implementation. For example, one might define an OrderingMixin that implements several ordering operators in terms of other operators.
class OrderingMixin: def __ge__(self, other): return self > other or self == other def __le__(self, other): return self < other or self == other def __ne__(self, other): return not self == other # ...and so on
Using this abstract base class - more properly, a concrete mixin - allows a programmer to define a limited set of operators and let the mixin in effect "derive" the others.
By combining these two orthogonal systems, we're able to both a) provide functionality, and b) alert consumer systems to the presence and availability of this functionality. For example, since the OrderingMixin class above satisfies the interface and semantics expressed in the Ordering role, we say the mixin performs the role:
@perform_role(Ordering) class OrderingMixin: def __ge__(self, other): return self > other or self == other def __le__(self, other): return self < other or self == other def __ne__(self, other): return not self == other # ...and so on
Now, any class that uses the mixin will automatically -- that is, without further programmer effort -- be tagged as performing the Ordering role.
The separation of concerns into two distinct, orthogonal systems is desirable because it allows us to use each one separately. Take, for example, a third-party package providing a RecursiveHash role that indicates a container takes its contents into account when determining its hash value. Since Python's built-in tuple and frozenset classes follow this semantic, the RecursiveHash role can be applied to them.
>>> perform_role(RecursiveHash)(tuple) >>> perform_role(RecursiveHash)(frozenset)
Now, any code that consumes RecursiveHash objects will now be able to consume tuples and frozensets.
Perl 6 allows instances to perform different roles than their class. These changes are local to the single instance and do not affect other instances of the class. For example:
my_elf = Elf() my_elf.goes_on_quest() my_elf.becomes_evil() now_performs(my_elf, Thieving) # Only this one elf is a thief my_elf.steals(["purses", "candy", "kisses"])
In Perl 6, this is done by creating an anonymous class that inherits from the instance's original parent and performs the additional role(s). This is possible in Python 3, though whether it is desirable is still is another matter.
Inclusion of this feature would, of course, make it much easier to express the works of Charles Dickens in Python:
>>> from literature import role, BildungsRoman >>> from dickens import Urchin, Gentleman >>> >>> with BildungsRoman() as OliverTwist: ... mr_brownlow = Gentleman() ... oliver, artful_dodger = Urchin(), Urchin() ... now_performs(artful_dodger, [role.Thief, role.Scoundrel]) ... ... oliver.has_adventures_with(ArtfulDodger) ... mr_brownlow.adopt_orphan(oliver) ... now_performs(oliver, role.RichWard)
Neal Norwitz has requested the ability to make assertions about the presence of attributes using the same mechanism used to require methods. Since roles take effect at class definition-time, and since the vast majority of attributes are defined at runtime by a class's __init__() method, there doesn't seem to be a good way to check for attributes at the same time as methods.
It may still be desirable to include non-enforced attributes in the role definition, if only for documentation purposes.
Under the proposed semantics, it is possible for roles to have roles of their own.
@perform_role(Y) class X(metaclass=Role): ...
While this is possible, it is meaningless, since roles are generally not instantiated. There has been some off-line discussion about giving meaning to this expression, but so far no good ideas have emerged.
It is currently not possible to ask a class if its instances perform a given role. It may be desirable to provide an analogue to performs() such that
>>> isinstance(my_dwarf, Dwarf) True >>> performs(my_dwarf, Surly) True >>> performs(Dwarf, Surly) False >>> class_performs(Dwarf, Surly) True
An early draft of this PEP included a separate mechanism for dynamically assigning a role to a class. This was spelled
>>> now_perform(Dwarf, GoldMiner)
This same functionality already exists by unpacking the syntactic sugar provided by decorators:
At issue is whether dynamic role assignment is sufficiently important to warrant a dedicated spelling.
Though the phrasings laid out in this PEP are designed so that the roles system could be shipped as a stand-alone package, it may be desirable to add special syntax for defining, assigning and querying roles. One example might be a role keyword, which would translate
class MyRole(metaclass=Role): ...
role MyRole: ...
Assigning a role could take advantage of the class definition arguments proposed in PEP 3115:
class MyClass(performs=MyRole): ...
A reference implementation is forthcoming.
Thanks to Jeffery Yasskin, Talin and Guido van Rossum for several hours of in-person discussion to iron out the differences, overlap and finer points of roles and abstract base classes.
This document has been placed in the public domain.