metautils3 0.1.2

Template Model

Why do we need or want to write class templates.

Consider the two metaclasses.

class AllLower(type):
    def __new__(mcls, name, bases, dict_):
        dict_ = {k.lower(): v for k, v in dict_.items()}
        return super().__new__(mcls, name, bases, dict_)


class MethodCatcher(type):
    def __new__(mcls, name, bases, dict_):
        dict_['methods'] = [v for v in dict_.values() if callable(v)]
        return super().__new__(mcls, name, bases, dict_)

What would we do if we wanted to make a class that used BOTH of these metaclasses? Using a class that subclasses both AllLower and MethodCatcher does not work, what we want is a way to chain them.

With the class template model, we could have written our metaclasses as:

from metautils3 import T, templated

class AllLower(T):
    @templated
    def __new__(mcls, name, bases, dict_):
        dict_ = {k.lower(): v for k, v in dict_.items()}
        return super().__new__(mcls, name, bases, dict_)


class MethodCatcher(T):
    @templated
    def __new__(mcls, name, bases, dict_):
        dict_['methods'] = [v for v in dict_.values() if callable(v)];
        return super().__new__(mcls, name, bases, dict_)

Python 2 style super calls will also work, like: super(AllLower, mcls). We can write the above classes without using super by delagating to T just like we would delegate to a concrete class, for example:

from metautils3 import T, templated

 class AllLower(T):
     @templated
     def __new__(mcls, name, bases, dict_):
         dict_ = {k.lower(): v for k, v in dict_.items()}
         return T.__new__(mcls, name, bases, dict_)


 class MethodCatcher(T):
     @templated
     def __new__(mcls, name, bases, dict_):
         dict_['methods'] = [v for v in dict_.values() if callable(v)];
         return T.__new__(mcls, name, bases, dict_)

Inside the context of a templated function, T with refer to the concrete class that was used to instantiate an instance of the template. Another name that will be changed out from under you is the class name itself. When you are in the context of a method, the class name will actually resolve to the concrete type.

Now we can define classes that use BOTH of these metaclasses like so:

class C(object, metaclass=MethodCatcher(AllLower())):
    def F():
        pass

    def g():
        pass

    a = 'a'
    B = 'b'

We can see that this applied the composition of the metaclasses.

>>> C.f
<function __main__.C.F>
>>> C.g
<function __main__.C.g>
>>> C.b
'b'
>>> C.a
'a'
>>> C.methods
[<function __main__.C.g>, <function __main__.C.F>]

The order that the metaclasses are composed is explicit as they act as transformers over each other.

Template

While the previous example only showed metaclasses, you can use this for any class; however, it is most useful for metaclasses where having a compatible metaclass hierarchy is important.

A Template is a callable that takes a type object and returns a new type object. It takes the following arguments:

• base: A type object. default: type.

• adjust_name: Should we prepend the name of the base to the new type object. default: True.

These can be chained together with any concrete metaclass at the end, e.g.:

new_class = m(n,p(q(...z(type)...)))

You can also use the compose function to do this:

from metautils3 import compose

new_class_template = compose(m, n, p, q, ..., z)

Differences with metautils

metautils3 uses far more experimental features, including bytecode and code object transformations that allow for more work to be done implicitly. This is how the T object can refernece the template argument inside of a method, or how we can get super to work as intended. This package also calls into ctypes and other CPython specific code, making it less portable and more difficult to maintain. This is mainly an interesting proof of concept to push metautils to the limits. For any production code, I have to recommend you use the more stable version.