Using a class' __new__ method as a Factory: __init__ gets called twice

Question

I encountered a strange bug in python where using the __new__ method of a class as a factory would lead to the __init__ method of the instantiated class to be called twice.

The idea was originally to use the __new__ method of the mother class to return a specific instance of one of her children depending on the parameters that are passed, without having to declare a factory function outside of the class.

I know that using a factory function would be the best design-pattern to use here, but changing the design pattern at this point of the project would be costly. My question hence is: is there a way to avoid the double call to __init__ and get only a single call to __init__ in such a schema ?

class Shape(object):
    def __new__(cls, desc):
        if cls is Shape:
            if desc == 'big':   return Rectangle(desc)
            if desc == 'small': return Triangle(desc)
        else:
            return super(Shape, cls).__new__(cls, desc)

    def __init__(self, desc):
        print "init called"
        self.desc = desc

class Triangle(Shape):
    @property
    def number_of_edges(self): return 3

class Rectangle(Shape):
    @property
    def number_of_edges(self): return 4

instance = Shape('small')
print instance.number_of_edges

>>> init called
>>> init called
>>> 3

Any help greatly appreciated.

Solution 1

When you construct an object Python calls its __new__ method to create the object then calls __init__ on the object that is returned. When you create the object from inside __new__ by calling Triangle() that will result in further calls to __new__ and __init__.

What you should do is:

class Shape(object):
    def __new__(cls, desc):
        if cls is Shape:
            if desc == 'big':   return super(Shape, cls).__new__(Rectangle)
            if desc == 'small': return super(Shape, cls).__new__(Triangle)
        else:
            return super(Shape, cls).__new__(cls, desc)

which will create a Rectangle or Triangle without triggering a call to __init__ and then __init__ is called only once.

Edit to answer @Adrian's question about how super works:

super(Shape,cls) searches cls.__mro__ to find Shape and then searches down the remainder of the sequence to find the attribute.

Triangle.__mro__ is (Triangle, Shape, object) and Rectangle.__mro__ is (Rectangle, Shape, object) while Shape.__mro__ is just (Shape, object). For any of those cases when you call super(Shape, cls) it ignores everything in the mro squence up to and including Shape so the only thing left is the single element tuple (object,) and that is used to find the desired attribute.

This would get more complicated if you had a diamond inheritance:

class A(object): pass
class B(A): pass
class C(A): pass
class D(B,C): pass

now a method in B might use super(B, cls) and if it were a B instance would search (A, object) but if you had a D instance the same call in B would search (C, A, object) because the D.__mro__ is (B, C, A, object).

So in this particular case you could define a new mixin class that modifies the construction behaviour of the shapes and you could have specialised triangles and rectangles inheriting from the existing ones but constructed differently.

Solution 2

After posting my question, I continued searching for a solution an found a way to solve the problem that looks like a bit of a hack. It is inferior to Duncan's solution, but I thought it could be interesting to mention none the less. The Shapeclass becomes:

class ShapeFactory(type):
    def __call__(cls, desc):
        if cls is Shape:
            if desc == 'big':   return Rectangle(desc)
            if desc == 'small': return Triangle(desc)
        return type.__call__(cls, desc)

class Shape(object):
    __metaclass__ = ShapeFactory 
    def __init__(self, desc):
        print "init called"
        self.desc = desc

Comments

[ncoghlan]

This is not correct - the first __init__ call happens inside the outer __new__ call (when Triangle() and Rectangle() are called), but then, because an instance of Shape is returned by __new__, the original Shape() call invokes __init__ again on that already initialised object. Note that if the object returned by __new__() is not an instance of Shape() then __init__ won't be called (which can foul up attempts to observe this behaviour if the class hierarchy isn't right).

[xApple]

Indeed, both are syntactically identical. My reluctance stems from the fact that if I define a function called "Shape", I must rename my class to something like "_Shape". This will cause some variable renaming, of course, but mostly it will have complicated consequences on other things like the documentation that is generated by sphinx-autodoc.

[kindall]

I suppose you do want to expose documentation on those classes so you can't just declare them in the function. You might try just reassigning the instance's __class__ attribute in __init__() and not messing with __new__() at all.

[kindall]

... one of the pitfalls of this, by the way, is that you will have to manually re-bind the instance methods from the correct class to the instance.

[kindall]

Another approach might be to use a metaclass to override the __call__() method on the class, so that the () syntax does not necessarily instantiate the class.

[kindall]

... in fact, the metaclass approach does work, but the solution you accepted is much simpler.

[Stefan]

Wouldn't it be better to use return Rectangle.__new__(Rectangle) because this would guarantee that __new__ of Rectangle gets called if it's defined?

[Duncan]

@Georg, If you do that you are going to have to be pretty careful to avoid infinite recursion. Any class specific initialisation should be in __init__ so I think it is pretty safe here to assume that __new__'s only job is to create an object of the correct type.

[martineau]

The code in your answer fixes the problem but doesn't explain why Shape.__init__() was called twice in the OP's code, even when Shape.__new__() returns a Triangle(desc). This seems contrary to what the docs say: "If __new__() does not return an instance of cls, then the new instance’s __init__() method will not be invoked."

[martineau]

Never mind, I'm guessing in the OP's code it's isinstance(<returned object>, Shape) being True for a Triangle object that's allowing the Shape.__init__() method to be called on it.

[javawizard]

@martineau: Precisely, __init__ is called if the returned object is an instance of the specified class or a subclass. (It's my frank opinion that this is a Bad Thing, but alas, it's the way it is.)

[Andy Hayden]

Why do you say this is inferior to Duncan's solution. This seems much clearer to what's going on, imo less hacky. Also meta.

[xApple]

I thought it was less obvious because it adds a fourth class to the program and involves the __metaclass__ black magic that not everyone is familiar with.

[Bob]

@GeorgSchölly super(Shape, cls) - in general, how does super use the first and second arguments to return the thing it returns? I know cls must be a subclass of Shape but how does the thing that super returns relate to Shape and cls?

[Stefan]

@Adrian: I don't know all the details of super by heart, but you can have a look at the documentation or the code. It must be something like this: super(Class, instance).method(args) is the same as Class.method(instance, args) where instance is usually called self inside the method.

[Duncan]

@Adrian, I added a description of how super works.

[martineau]

Although most folks probably don't care, one advantage of using inheritance instead of a metaclass is that the syntax for specifying metaclasses is different in Python 3 that it is in Python 2—on the other hand, inheritance is written similarly in both of them. That fact could make it a better fit when attempting to write code that will work unchanged in both versions of Python.

[geekscrap]

What would be the equivalent in python 3?

[Marine Galantin]

IMHO having metaclass is cleaner because here, one is trying to implemented a top down inheritence (creating child from parents) instead of the normal inheritance order down-top (creating parents from child). Metaclass seems to be the right way to do it :) Also, the whole purpose of calland new is to deal with such cases...