Using the __call__ method of a metaclass instead of __new__?

It’s a matter of lifecycle phases and what you have access to. __call__ gets called after __new__ and is passed the initialization parameters before they get passed on to __init__, so you can manipulate them. Try this code and study its output:

class Meta(type):
    def __new__(cls, name, bases, newattrs):
        print "new: %r %r %r %r" % (cls, name, bases, newattrs,)
        return super(Meta, cls).__new__(cls, name, bases, newattrs)

    def __call__(self, *args, **kw):
        print "call: %r %r %r" % (self, args, kw)
        return super(Meta, self).__call__(*args, **kw)

class Foo:
    __metaclass__ = Meta

    def __init__(self, *args, **kw):
        print "init: %r %r %r" % (self, args, kw)

f = Foo('bar')
print "main: %r" % f

One difference is that by defining a metaclass __call__ method you are demanding that it gets called before any of the class’s or subclasses’s __new__ methods get an opportunity to be called.

class MetaFoo(type):
    def __call__(cls,*args,**kwargs):
        print('MetaFoo: {c},{a},{k}'.format(c=cls,a=args,k=kwargs))

class Foo(object):
    __metaclass__=MetaFoo

class SubFoo(Foo):
    def __new__(self,*args,**kwargs):
        # This never gets called
        print('Foo.__new__: {a},{k}'.format(a=args,k=kwargs))

 sub=SubFoo()
 foo=Foo()

 # MetaFoo: <class '__main__.SubFoo'>, (),{}
 # MetaFoo: <class '__main__.Foo'>, (),{}

Notice that SubFoo.__new__ never gets called. In contrast, if you define Foo.__new__ without a metaclass, you allow subclasses to override Foo.__new__.

Of course, you could define MetaFoo.__call__ to call cls.__new__, but that’s up to you. By refusing to do so, you can prevent subclasses from having their __new__ method called.

I don’t see a compelling advantage to using a metaclass here. And since “Simple is better than complex”, I’d recommend using __new__.

The subtle differences become a bit more visible when you carefully observe the execution order of these methods.

class Meta_1(type):
    def __call__(cls, *a, **kw):
        print "entering Meta_1.__call__()"
        rv = super(Meta_1, cls).__call__(*a, **kw)
        print "exiting Meta_1.__call__()"
        return rv

class Class_1(object):
    __metaclass__ = Meta_1
    def __new__(cls, *a, **kw):
        print "entering Class_1.__new__()"
        rv = super(Class_1, cls).__new__(cls, *a, **kw)
        print "exiting Class_1.__new__()"
        return rv

    def __init__(self, *a, **kw):
        print "executing Class_1.__init__()"
        super(Class_1,self).__init__(*a, **kw)

Note that the code above doesn’t actually do anything other than log what we’re doing. Each method defers to its parent implementation i.e. its default. So beside logging it’s effectively as if you had simply declared things as follows:

class Meta_1(type): pass
class Class_1(object):
    __metaclass__ = Meta_1

And now let’s create an instance of Class_1

c = Class_1()
# entering Meta_1.__call__()
# entering Class_1.__new__()
# exiting Class_1.__new__()
# executing Class_1.__init__()
# exiting Meta_1.__call__()

Therefore if type is the parent of Meta_1 we can imagine a pseudo implementation of type.__call__() as such:

class type:
    def __call__(cls, *args, **kwarg):

        # ... a few things could possibly be done to cls here... maybe... or maybe not...

        # then we call cls.__new__() to get a new object
        obj = cls.__new__(cls, *args, **kwargs)

        # ... a few things done to obj here... maybe... or not...

        # then we call obj.__init__()
        obj.__init__(*args, **kwargs)

        # ... maybe a few more things done to obj here

        # then we return obj
        return obj

Notice from the call order above that Meta_1.__call__() (or in this case type.__call__()) is given the opportunity to influence whether or not calls to Class_1.__new__() and Class_1.__init__() are eventually made. Over the course of its execution Meta_1.__call__() could return an object that hasn’t even been touched by either. Take for example this approach to the singleton pattern:

class Meta_2(type):
    __Class_2_singleton__ = None
    def __call__(cls, *a, **kw):
        # if the singleton isn't present, create and register it
        if not Meta_2.__Class_2_singleton__:
            print "entering Meta_2.__call__()"
            Meta_2.__Class_2_singleton__ = super(Meta_2, cls).__call__(*a, **kw)
            print "exiting Meta_2.__call__()"
        else:
            print ("Class_2 singleton returning from Meta_2.__call__(), "
                    "super(Meta_2, cls).__call__() skipped")
        # return singleton instance
        return Meta_2.__Class_2_singleton__

class Class_2(object):
    __metaclass__ = Meta_2
    def __new__(cls, *a, **kw):
        print "entering Class_2.__new__()"
        rv = super(Class_2, cls).__new__(cls, *a, **kw)
        print "exiting Class_2.__new__()"
        return rv

    def __init__(self, *a, **kw):
        print "executing Class_2.__init__()"
        super(Class_2, self).__init__(*a, **kw)

Let’s observe what happens when repeatedly trying to create an object of type Class_2

a = Class_2()
# entering Meta_2.__call__()
# entering Class_2.__new__()
# exiting Class_2.__new__()
# executing Class_2.__init__()
# exiting Meta_2.__call__()

b = Class_2()
# Class_2 singleton returning from Meta_2.__call__(), super(Meta_2, cls).__call__() skipped

c = Class_2()
# Class_2 singleton returning from Meta_2.__call__(), super(Meta_2, cls).__call__() skipped

print a is b is c
True

Now observe this implementation using a class’ __new__() method to try to accomplish the same thing.

import random
class Class_3(object):

    __Class_3_singleton__ = None

    def __new__(cls, *a, **kw):
        # if singleton not present create and save it
        if not Class_3.__Class_3_singleton__:
            print "entering Class_3.__new__()"
            Class_3.__Class_3_singleton__ = rv = super(Class_3, cls).__new__(cls, *a, **kw)
            rv.random1 = random.random()
            rv.random2 = random.random()
            print "exiting Class_3.__new__()"
        else:
            print ("Class_3 singleton returning from Class_3.__new__(), "
                   "super(Class_3, cls).__new__() skipped")

        return Class_3.__Class_3_singleton__ 

    def __init__(self, *a, **kw):
        print "executing Class_3.__init__()"
        print "random1 is still {random1}".format(random1=self.random1)
        # unfortunately if self.__init__() has some property altering actions
        # they will affect our singleton each time we try to create an instance 
        self.random2 = random.random()
        print "random2 is now {random2}".format(random2=self.random2)
        super(Class_3, self).__init__(*a, **kw)

Notice that the above implementation even though successfully registering a singleton on the class, does not prevent __init__() from being called, this happens implicitly in type.__call__() (type being the default metaclass if none is specified). This could lead to some undesired effects:

a = Class_3()
# entering Class_3.__new__()
# exiting Class_3.__new__()
# executing Class_3.__init__()
# random1 is still 0.282724600824
# random2 is now 0.739298365475

b = Class_3()
# Class_3 singleton returning from Class_3.__new__(), super(Class_3, cls).__new__() skipped
# executing Class_3.__init__()
# random1 is still 0.282724600824
# random2 is now 0.247361634396

c = Class_3()
# Class_3 singleton returning from Class_3.__new__(), super(Class_3, cls).__new__() skipped
# executing Class_3.__init__()
# random1 is still 0.282724600824
# random2 is now 0.436144427555

d = Class_3()
# Class_3 singleton returning from Class_3.__new__(), super(Class_3, cls).__new__() skipped
# executing Class_3.__init__()
# random1 is still 0.282724600824
# random2 is now 0.167298405242

print a is b is c is d
# True

I thought a fleshed out Python 3 version of pyroscope’s answer might be handy for someone to copy, paste and hack about with (probably me, when I find myself back at this page looking it up again in 6 months). It is taken from this article:

class Meta(type):

     @classmethod
     def __prepare__(mcs, name, bases, **kwargs):
         print('  Meta.__prepare__(mcs=%s, name=%r, bases=%s, **%s)' % (
             mcs, name, bases, kwargs
         ))
         return {}

     def __new__(mcs, name, bases, attrs, **kwargs):
         print('  Meta.__new__(mcs=%s, name=%r, bases=%s, attrs=[%s], **%s)' % (
             mcs, name, bases, ', '.join(attrs), kwargs
         ))
         return super().__new__(mcs, name, bases, attrs)

     def __init__(cls, name, bases, attrs, **kwargs):
         print('  Meta.__init__(cls=%s, name=%r, bases=%s, attrs=[%s], **%s)' % (
             cls, name, bases, ', '.join(attrs), kwargs
         ))
         super().__init__(name, bases, attrs)

     def __call__(cls, *args, **kwargs):
         print('  Meta.__call__(cls=%s, args=%s, kwargs=%s)' % (
             cls, args, kwargs
         ))
         return super().__call__(*args, **kwargs)

print('** Meta class declared')

class Class(metaclass=Meta, extra=1):

     def __new__(cls, myarg):
         print('  Class.__new__(cls=%s, myarg=%s)' % (
             cls, myarg
         ))
         return super().__new__(cls)

     def __init__(self, myarg):
         print('  Class.__init__(self=%s, myarg=%s)' % (
             self, myarg
         ))
         self.myarg = myarg
         super().__init__()

     def __str__(self):
         return "<instance of Class; myargs=%s>" % (
             getattr(self, 'myarg', 'MISSING'),
         )

print('** Class declared')

Class(1)
print('** Class instantiated')

Outputs:

** Meta class declared
  Meta.__prepare__(mcs=<class '__main__.Meta'>, name='Class', bases=(), **{'extra': 1})
  Meta.__new__(mcs=<class '__main__.Meta'>, name='Class', bases=(), attrs=[__module__, __qualname__, __new__, __init__, __str__, __classcell__], **{'extra': 1})
  Meta.__init__(cls=<class '__main__.Class'>, name='Class', bases=(), attrs=[__module__, __qualname__, __new__, __init__, __str__, __classcell__], **{'extra': 1})
** Class declared
  Meta.__call__(cls=<class '__main__.Class'>, args=(1,), kwargs={})
  Class.__new__(cls=<class '__main__.Class'>, myarg=1)
  Class.__init__(self=<instance of Class; myargs=MISSING>, myarg=1)
** Class instantiated

Another great resource highlighted by the same article is David Beazley’s PyCon 2013 Python 3 Metaprogramming tutorial.

In the particular example given in the question, overriding __call__ in the metaclass is just superior to overriding __new__ in the class.

suppose I want to have custom behavior when the class is called, for example caching instead of creating fresh objects

1. If the purpose of caching is efficiency, then caching __new__‘s result is not optimal because __init__ is executed anyhow (Data Model: Basic Customization). For example:

   from functools import lru_cache
   
   class MyClass:
   
       @lru_cache(maxsize=None)
       def __new__(cls, *args, **kwargs):
           return super().__new__(cls)
   
       def __init__(self, ...)
           "Always executed. Even on cache hit."
   

2. Caching __new__‘s result is sound only if __init__ has no noticeable effect on an already initialized instance. Otherwise the execution of __init__ on a cached object could lead to annoying side-effects on its other references.

3. Caching at the metaclass level avoids both the performance issue of 1. and the correctness issue of 2. For example:

   from functools import lru_cache
   
   class CachedInstances(type):
   
       @lru_cache(maxsize=None)
       def __call__(cls, *args, **kwargs):
           return super().__call__(*args, **kwargs)
   
   class MyClass(metaclass=CachedInstances):
   
       def __init__(self, ...)
           "Only executed on cache miss."
   

Note that Michael Ekoka’s answer already mention undesired effects that may arise due to repeated __init__ execution within the override __new__ approach (as in my item 2).