How do you decide between using decorators and inheritance when both are possible?
E.g., this problem has two solutions.
I'm particularly interested in Python.
Decorators...:
...should be used if what you are trying to do is "wrapping". Wrapping consists of taking something, modifying (or registering it with something), and/or returning a proxy object that behaves "almost exactly" like the original.
...are okay for applying mixin-like behavior, as long as you aren't creating a large stack of proxy objects.
...have an implied "stack" abstraction:
e.g.
#decoA
#decoB
#decoC
def myFunc(...): ...
...
Is equivalent to:
def myFunc(...): ...
...
myFunc = decoA(decoB(decoC(myFunc))) #note the *ordering*
Multiple inheritance...:
... is best for adding methods to classes; you cannot use it to decorate functions easily. In this context, it can be used to achieve mixin-like behavior if all you need is a set of "duck-typing style" extra methods.
... may be a bit unwieldy if your problem is not a good match for it, with issues with superclass constructors, etc. For example, the subclasses __init__ method will not be called unless it is called explicitly (via the method-resolution-order protocol)!
To sum up, I would use decorators for mixin-like behavior if they didn't return proxy objects. Some examples would include any decorator which returns the original function, slightly modified (or after registering it somewhere or adding it to some collection).
Things you will often find decorators for (like memoization) are also good candidates, but should be used in moderation if they return proxy objects; the order they are applied matter. And too many decorators on top of one another is using them in a way they aren't intended to be used.
I would consider using inheritance if it was a "classic inheritance problem", or if all I needed for the mixin behavior were methods. A classic inheritance problem is one where you can use the child wherever you could use the parent.
In general, I try to write code where it is not necessary to enhance arbitrary things.
The problem you reference is not deciding between decorators and classes. It is using decorators, but you have the option of using either:
a decorator, which returns a class
a decorator, which returns a function
A decorator is just a fancy name for the "wrapper" pattern, i.e. replacing something with something else. The implementation is up to you (class or function).
When deciding between them, it's completely a matter of personal preference. You can do everything you can do in one with the other.
if decorating a function, you may prefer decorators which return proxy functions
if decorating a class, you may prefer decorators which return proxy classes
(Why is it a good idea? There may be assumptions that a decorated function is still a function, and a decorated class is still a class.)
Even better in both cases would be to use a decorator which just returns the original, modified somehow.
edit: After better understanding your question, I have posted another solution at Python functools.wraps equivalent for classes
The other answers are quite great, but I wanted to give a succinct list of pros and cons.
The main advantage of mixins is that the type can be checked at runtime using isinstance and it can be checked with linters like MyPy. Like all inheritance, it should be used when you have an is-a relationship. For example dataclass should probably have been a mixin in order to expose dataclass-specific introspection variables like the list of dataclass fields.
Decorators should be preferred when you don't have an is-a relationship. For example, a decorator that propagates documentation from another class, or registers a class in some collection.
Decoration typically only affects the class it decorates, but not classes that inherit from the base class:
#decorator
class A:
... # Can be affected by the decorator.
class B(A):
... # Not affected by the decorator in most cases.
Now that Python has __init_subclass__, everything that decorators can do can be done with mixins, and they typically do affect child subclasses:
class A(Mixin):
... # Is affected by Mixin.__init_subclass__.
class B(A):
... # Is affected by Mixin.__init_subclass__.
Mixins have another advantage, which is that they can provide empty base class methods. Child classes can override these methods with some "augmenting" behavior, and then call super. The decorator cannot easily provide such base class methods. This is another way in which mixins are more flexible.
In summary, the questions you should ask when deciding between a mixin and decoration are:
Is there an is-a pattern?
Would you ever call isinstance?
Would you use the mixin in a type annotation?
Do you want the behavior to affect child classes?
Do you need augmenting methods?
In general, lean towards inheritance.
If both are equivalent, I would prefer decorators, since you can use the same decorator for many classes, while inheriting apply to only one specific class.
Personally, I would think in terms of code reuse. Decorator is sometimes more flexible than inheritance.
Let's take caching as an example. If you want to add caching facility to two classes in your system: A and B, with inheritance, you'll probably wind up having ACached and BCached. And by overriding some of the methods in these classes, you'll probably duplicate a lot of codes for the same caching logic. But if you use decorator in this case, you only need to define one decorator to decorate both classes.
So, when deciding which one to use, you may first want to check if the extended functionality is only specific to this class or if the same extended functionality can be reused in other parts of your system. If it cannot be reused, then inheritance should probably do the job. Otherwise, you can think about using decorator.
Related
Functions in a python class can be either instance methods, class methods or static methods.
The former is characterised by the self as its first (implicit) argument, acts directly on the instance of the class, and does not require any decorators to be treated as such.
The other two, however, need decorators #classmethod and #staticmethod before the name of the method - this is why I refer to the instance method as the "default" one, i.e. the one for which a wrapper is not needed.
My question is: suppose I am in a class, and I am breaking up my calculation into several functions for readibility. Only one of these methods will need access to the self.something variables that I share instance-wise, but most of the others do not need to know about the class they belong to - they are just there for "housekeeping".
Should make these functions (the ones that do not need any self.something knowledge) all #staticmethod? Doing so would require a decorator and hence an extra step. It would be easier (not requiring the extra step of using a decotrator) for every method to just be an instance method, thus inheritig a lot of potential but also waisting it since it is not needed for the scope of the functions in question.
Why is the instance method the "default"? Why not have every method a static method by default, and give it the extra functionality associated with being a instance method with a wrapper?
The reason to default to instance methods is because that's usually what you want when you're doing object oriented programming. I can't think of a single language that claims to support OOP and has methods default to anything but instance methods. Classes are templates for "data with behaviors", so the default is to make methods that provide behaviors to each instantiation of the class. If you just want a collection of functions, you can just define them at the top level of a module and save the unnecessary class after all.
In general, #staticmethod is used to mean "I know this isn't a behavior of the class or its instances, but it helps implement the real behaviors and isn't very useful outside the class, so I'll namespace it inside it." If the features are useful outside the class, you'd just make it a plain top-level function rather putting it inside the class at all. It is advantageous to use #staticmethod where appropriate; it's a little faster to call than an instance method, so if you don't need the instance, #staticmethod will speed up your code a bit (note: This may not be true in 3.7+, where they added an optimization to avoid the creation of bound methods, which may speed up instance/class methods).
#classmethod basically has two use cases:
(Primary) Defining alternate constructors in a subclass friendly way (the cls it receives is the actual subclass, if applicable, not just the class it was defined in)
(Mostly unnecessary) As an alternative to #staticmethod when the method needs to call other static methods and you'd rather not have to refer to the class by name over and over
Point is, #staticmethod is mostly for when you're opting out of OOP, and #classmethods are for niche use cases; instance methods are just more useful, so they're the default. Beyond that, as a historical note, static and class methods were introduced later, so making them the default would have broken all existing Python code, for no real benefit.
The main reason to use #staticmethod over instance methods with an ignored self (when self isn't needed) is that it will continue to work when called on the class itself, not just on instances of the class; if you tried to call MyClass.notreallystatic(), it would die for lack of a self, while MyClass.actuallystatic() would work.
I have a mix-in class called WithAutoNumbering for classes that need a special numbering of a given attribute. Appart from that I have a nice class mix-in called WithIndexing for those classes that need indexing capabilities... which needs the capabilities of WithAutoNumbering.
Some classes need numbering but not indexing, so mixing them together is not a good idea.
The dilemma is, should WithIndexing inherit from WithAutoNumbering? or each class that needs WithIndexing should also inherit from WithAutoNumbering as well?
I.e. this, with CoolClass being the one that has to implement indexing:
class WithAutoNumbering(object):
...
class WithIndexing(WithAutoNumbering):
...
class CoolClass(WithIndexing):
...
or this
class WithAutoNumbering(object):
...
class WithIndexing(object):
...
class CoolClass(WithIndexing, WithAutoNumbering):
...
On the one hand, the first approach is more succint, and makes sure that you can't try to use WithIndexing withouth WithAutoNumbering. On the other hand, I have always read (and found it agreeable) that mix-ins should not have hierarchy, i.e. inherit only from object, in order to avoid spaghettization of the whole class hierarchy with ununderstandable __mro__s
The issue with making a choice about what your mixins inherit from is that your choice will affect the final MRO of classes which use those mixins.
should WithIndexing inherit from WithAutoNumbering
As you say, WithIndexing uses WithAutoNumbering. It's a funny kind of class which cannot be used on its own at all; and by inheriting from WithAutoNumbering, WithIndexing can override WithAutoNumbering members.
In general, it will be easier for each level to override methods found above, the more you set this up as single inheritance.
However, if you choose to design the other way (such that you need to inherit from both to make a class capable of using WithIndexing), you will likely be completely fine, until something overrides a method found in WithAutoNumbering. In that case, the order in which the classes appear in the base list may affect their order of the MRO, in which case you may have surprising results. If you need the power to affect the MRO in that way, you should have each of these mixins inherit from object. You probably don't need that, though.
The Liskov substitution principle applies. Would a class designed to work with WithAutoNumbering work with WithIndexing instead (whether or not it actually uses or needs whatever WithIndexing adds)? If not, then WithIndexing should just extend object.
I understand mixin as what looks like inheritance but what is more like composition.
(edit: I tend to think giving additional functionality/attributes by mixin rather than giving another is-a relationship.)
Mentally, I'm saying something like this when I use mixin: I'm giving you this mixin you are missing, rather than you are actually this mixin-type as well.(is-a)
And I read few times, you should prefer composition over inheritance.
We could just use straight compositions instead of mixins, what is mixin for?
If I have to guess, it's because my_instance.foo() is easier than my_instance.another_instance.foo()?
(You can use my_instance.foo() if mixin has foo(), you need my_instance.another_instance.foo() when you composite another_instance as an attribute of my_instance)
Are there any other reason?
Edit:
So even though I feel it's has-a, mixin is still is-a relationship. and benefit you get when you use is-a here is, cleaner interface. That' how I interpret delnan's answer.
(Since mixin doesn't give you is-a relationship, it gives you has-a)
Wrong, it does give you a is-a relationship. Consider class C(A, B). issubclass(C, A) is true and so is issubclass(C, B). The call C().method_of_B() calls the method of B if it isn't overridden somewhere in the MRO (which can happen with single inheritance too!), and similarly for methods of A. The attribute access C().attr_of_{a,b} gives you an attribute of A or B too (same caveat as with methods).
And I read few times, you should prefer composition over inheritance.
This rule of thumb exists because many people tend to use inheritance where it isn't appropriate, not because inheritance isn't a useful tool in any case. This applies to Python too. Common (good) reasons include:
A is-a relationship would be wrong, for example because it breaks the Liskov Substitution Principle. This reason is not affected by mixins, see above.
Composition is easier or less error-prone, for various reasons. This reason is valid when mixins are possible too, if anything it's amplified as it can be a bit tricky (though not impossible) to get the MRO and sequence of super() calls right.
If I have to guess, it's because my_instance.foo() is easier than my_instance.another_instance.foo()?
That's just a few additional characters. More important is that you have to re-type wrapper methods for every single method you want to re-export (which would be all of them if you're emulating an is-a relationship via composition).
More conceptually, this manual work seems pointless and error-prone if C really is-a A. Mixins/multiple inheritance support the case where an object is more than what can or should be represented by a single bass class. For example, a UTF-8 string may be both a ByteBuffer and a UnicodeStream. In general, we have interfaces for this (or not; in Python this is implicit), but mixins also allow adding related functionality in a composable and convenient manner.
I once read (I think on a page from Microsoft) that it's a good way to use static classes, when you don't NEED two or more instances of a class.
I'm writing a program in Python. Is it a bad style, if I use #classmethod for every method of a class?
Generally, usage like this is better done by just using functions in a module, without a class at all.
It's terrible style, unless you actually need to access the class.
A static method [...] does not translate to a Python classmethod. Oh sure, it results in more or less the same effect, but the goal of a classmethod is actually to do something that's usually not even possible [...] (like inheriting a non-default constructor). The idiomatic translation of a [...] static method is usually a module-level function, not a classmethod or staticmethod.
source
In my experience creating a class is a very good solution for a number of reasons. One is that you wind up using the class as a 'normal' class (esp. making more than just one instance) more often than you might think. It's also a reasonable style choice to stick with classes for everthing; this can make it easier for others who read/maintain your code, esp if they are very OO - they will be comfortable with classes. As noted in other replies, it's also reasonable to just use 'bare' functions for the implementation. You may wish to start with a class and make it a singleton/Borg pattern (lots of examples if you googlefor these); it gives you the flexibility to (re)use the class to meet other needs. I would recommend against the 'static class' approach as being non-conventional and non-Pythonic, which makes it harder to read and maintain.
There are a few approaches you might take for this. As others have mentioned, you could just use module-level functions. In this case, the module itself is the namespace that holds them together. Another option, which can be useful if you need to keep track of state, is to define a class with normal methods (taking self), and then define a single global instance of it, and copy its instance methods to the module namespace. This is the approach taken by the standard library "random" module -- take a look at lib/python2.5/random.py in your python directory. At the bottom, it has something like this:
# Create one instance, seeded from current time, and export its methods
# as module-level functions. [...]
_inst = Random()
seed = _inst.seed
random = _inst.random
uniform = _inst.uniform
...
Or you can take the basic approach you described (though I would recommend using #staticmethod rather than #classmethod in most cases).
You might actually want a singleton class rather than a static class:
Making a singleton class in python
Or should I just explicitly reference the superclasses whose methods I want to call?
It seems brittle to repeat the names of super classes when referencing their constructors, but this page http://fuhm.net/super-harmful/ makes some good arguments against using super().
The book Expert Python Programming has discussed the topic of "super pitfalls" in chapter 3. It is worth reading. Below is the book's conclusion:
Super usage has to be consistent: In a class hierarchy, super should be used everywhere or nowhere. Mixing super and classic calls is a confusing practice. People tend to avoid super, for their code to be more explicit.
Edit: Today I read this part of the book again. I'll copy some more sentences, since super usage is tricky:
Avoid multiple inheritance in your code.
Be consistent with its usage and don't mix new-style and
old-style.
Check the class hierarchy before calling its methods in
your subclass.
You can use super, but as the article says, there are drawbacks. As long as you know them, there is no problem with using the feature. It's like people saying "use composition, not inheritance" or "never use global variables". If the feature exists, there is a reason. Just be sure to understand the why and the what and use them wisely.
super() tries to solve for you the problem of multiple inheritance; it's hard to replicate its semantics and you certainly shouldn't create any new semantics unless you're completely sure.
For single inheritance, there's really no difference between
class X(Y):
def func(self):
Y.func(self)
and
class X(Y):
def func(self):
super().func()
so I guess that's just the question of taste.
I like super() more because it allows you to change the inherited class (for example when you're refactoring and add an intermediate class) without changing it on all the methods.
The problem people have with super is more a problem of multiple inheritance. So it is a little unfair to blame super. Without super multiple inheritance is even worse. Michele Simionato nicely wrapped this up in his blog article on super:
On the other hand, one may wonder if
all super warts aren't hints of some
serious problem underlying. It may
well be that the problem is not with
super, nor with cooperative methods:
the problem may be with multiple
inheritance itself.
So the main lesson is that you should try to avoid multiple inheritance.
In the interest of consistency I always use super, even if for single inheritance it does not really matter (apart from the small advantage of not having to know the parent class name). In Python 3+ super is more convenient, so there one should definitely use super.
Yes, you should use super() over other methods. This is now the standard object inheritance model in Python 3.
Just stick to keyword arguments in your __init__ methods and you shouldn't have too many problems. Additionally you can use **kwargs to support additional parameters that are not defined in levels of the inheritance chain.
I agree that it is brittle, but no less so than using the name of the inherited class.