The Artima Developer Community
Sponsored Link

The Explorer
Mixins considered harmful/2
by Michele Simionato
January 6, 2009
Summary
In this post I clarify my position with respect to mixins, I show a few common design mistakes and I suggest some alternative solution.

Advertisement

A few conceptual issues with mixins

In the first article of this series I have discussed a very serious problem of the mixin approach, i.e. the namespace overpopulation issue.

http://www.phyast.pitt.edu/~micheles/python/Overpopulation.jpg

The namespace overpopulation issue

The overpopulation issue comes from the idea of growing functionality by adding more and more mixin classes, which is just plain wrong. It is true that you can use the idea in little frameworks with little damage, but that does not make it a good design solution. Small frameworks have a tendency to grow, and you should not start with a weak design.

Some reader argued that this is not a problem of mixins per se, but a problem of bad design. That is true, but I maintain that a technique which is so easy to misuse even by expert programmers, should be regarded with suspicion, especially when there are better solutions available. Moreover, I have a few conceptual issues with mixins - as implemented in most languages - which are independent of the overpopulation problem.

First of all, I think everybody agrees that the best way to solve a complex problem is to split it in smaller subproblems, by following the dividi et impera principle. The disturbing thing about mixins is that the principle is applied at the beginning (the problem is decomposed in smaller independent units) but at the end all the functionalities are added back to the client class as an undifferentiated soup of methods.

Therefore a design based on mixins looks clean to the framework writer - everything is well separated in his mind - but it looks messy to the framework user - she sees methods coming from all directions without a clear separation. It is really the same situation than using the from module import * idiom, which is rightly frowned upon.

I find it most unpythonic that mixins make the life of the framework writer easier, but the life of the framework reader more difficult, since the goal of Python is to make code easy to read, not easy to write. The scenario I have in mind is the usual one: a poor programmer who needs to debug an object coming from a gigantic framework which is terra incognita to her, without any documentation and with a strict deadline (do you see yourself in there?). In such conditions a framework heavily based on mixins makes things harder, since the programmer gets drowned under hundreds of methods which are properly ordered in mixin classes on the paper, but not on the battle field.

There is also another conceptual issue. The idea behind mixins is that they should be used for generic functionality which can be applied to different classes (think of mixins like Persistent, Comparable, Printable, etc.). But this is exactly the same situation where you want to use generic functions.

In this post of mine I actually argue that generic functions (a.k.a. multimethods) are a better solution than mixins. I also provide a very concrete example, which I think generalizes. The advantage of generic functions is that they are clearly defined outside classes, whereas the mixin approach is kind of schizophrenic: the functionality is actually defined externally, but that fact is made invisible to the final user.

I am a big fan of generic functions which are already used in the Python word - print is a generic function, the comparison operators are generic functions, numpy universal functions (ufunctions) are generic functions, etc - but should be used even more. With generic functions, mixins becomes useless. A side effect is that the class namespace becomes much slimmer: for instance, in CLOS classes are used just to contain state, whereas the methods live in a separate namespace. In most languages instead, classes are used as a namespace control mechanism, performing double duty - namespace control should be the job of modules.

A particularly bad usage of mixins

A tipical beginner's tutorial (for instance, I recommend Using Mix-ins with Python , by Chuck Esterbrook which is very well written and very informative, even if the point of view is exactly the opposite of mine) will tell you that mixins are used to add functionality to the classes they mix in. For instance a mixin class WithLog could be used to enhance a pre-existing class C with a logging capability:

class C(object):
  "A base class"
class WithLog(object):
  "A mixin class"
  @property
  def log(self):
    return logging.getLogger(self.__class__.__name__)
class C_WithLog(C, WithLog):
  "A mixin-enhanced class"

An example of usage is the following:

>>> c = C_WithLog()
>>> c.log.warn("hello")

That prints

WARNING:C_WithLog:hello.

The usage of mixins you see here is wrong: why would you use inheritance when you need just one method? You can just import the one method you need! Generally speaking, a mixin class has sense only when you have a set of methods which belong together: if you have a single method, or a set of disconnected methods, you are much better off by defining the methods externally, in an utility module, and then by importing them in the class namespace. Of course, here I am assuming that you really want the external method to end up in the class namespace, possibly because of interface requirements, but I am not saying that this is always a good idea. You can import the method in your class as simply as that:

class CWithLog(C):
  from utility import log # log is the property defined above

This approach is very little used in Python, probably because most people coming from other languages do not know it is possible, but it is in my opinion a much clearer solution than inheritance. The problem with inheritance is that it requires a substantial cognitive load: when I see the line of code class C_WithLog(C, WithLog) I immediately I ask myself many questions: which methods are exported by C_WithLog? is there any method of C which accidentally overrides one of the methods of C_WithLog? if yes, is there any method cooperation mechanism (super) or not? what are the ancestors of C_WithLog? which methods are coming from them? are such methods overridden by some C method? is there a cooperation mechanism on C_WithLog ancestors? What's the method resolution order of the hierarchy? On the other hand, if I see from utility import log I have very little to understand and very little to worry about. The only caution in this specific example is that I will have a single logger shared by all instances of the class since logging.getLogger(self.__class__.__name__) will return always the same object. If I need different loggers with different configurations for different instances I will have to override the .log attribute on a case by case basis, or I will have to use a different strategy, such as the dependency injection pattern, i.e. I will have to pass the logger to the constructor.

Acceptable usages of mixins

There are usages for mixins which are restricted in scope and not dangerous: for instance, you can use mixins for implementing the comparison interface, or the mapping interface. This is actually the approach suggested by the standard library, and by the new ABC's in Python 2.6. This is an acceptable usage: in this case there is no incontrollable growth of methods, since you are actually implementing well know interfaces - typically a few specific special methods. In order to give a practical example, let me discuss a toy application.

Suppose you want to define a PictureContainer class in an application to manage pictures and photos. A PictureContainer object may contain both plain pictures (instances of a Picture class) and PictureContainer objects, recursively. From the point of view of the Python programmer it could make sense to implement such a class by using a dictionary. A Picture object will contain information such as the picture title, the picture date, and a few methods to read and write the picture on the storage (the file system, a relation database, an object database like the ZODB or the AppEngine datastore, or anything else).

The first version of the PictureContainer class could be something like that:

class SimplePictureContainer(object):
    "A wrapper around the .data dictionary, labelled by an id"
    def __init__(self, id, pictures_or_containers):
      self.id = id
      self.data = {} # the inner dictionary
      for poc in pictures_or_containers:
        # both pictures and containers must have an .id
        self.data[poc.id] = poc

At this point, one realized that it is annoying to call the inner dictionary directly and that it would be nicer to expose its methods. A simple solution is to leverage on the standard library class UserDict.DictMixin which is there just for that use case. Since we are at it, we can also add the logging functionality: that means that the low-level interface (calling directly the inner dictionary methods) will not log whereas the high level interface will log:

class BetterPictureContainer(SimplePictureContainer, DictMixin):
  from utility import log

  def __getitem__(self, id):
    return self.data[id]

  def __setitem__(self, id, value):
    self.log.info('Adding or replacing %s into %s', id, self.id)
    self.data[id] = value

  def __delitem__(self, id):
    self.log.warn('Deleting %s', id)
    del self.data[id]

  def keys(self):
    return self.data.keys()

Using DictMixin is acceptable, since

  1. DictMixin provided to its subclasses the standard interface of a dictionary, a conceptually tied set of methods;
  2. the dictionary interface is well know to anybody knowing how to use dictionaries in Python, so that the cognitive load is zero;
  3. DictMixin allows a substantial code reuse: we redefined explicitly just 4 methods, but actually we are indirecty affecting 17 other methods: __cmp__, __contains__, __iter__, __len__, __repr__, clear, get, has_key, items, iteritems, iterkeys, itervalues, pop, popitem, setdefault, update, values: without DictMixin we would need to reimplement all of them.

However, notice that in this example the usage of DictMixin as mixin class is acceptable, but not optimal: the best solution is to use DictMixin as a base class, not as a mixin class.

The core problem is that we started from a wrong desing: we wrote SimplePictureContainer when we did not know of the existence of DictMixin. Now, a posteriori, we are trying to fix the mistake by using multiple inheritance, but that it not the Rigth Thing (TM) to do. The right thing would be to change the source code of SimplePictureContainer and to derive directly from DictMixin.

In the real world usually you do not have complete control of the code: you may leverage on a third party library with a design error, or simply with an old library, written when DictMixin did not exist. In such a situation you may have no way to modify the source code. Then using DictMixin and multiple inheritance is a perfectly acceptable workaround, but it is a workaround still, and it should not be traded for a clever design.

Moreover, even the best examples of mixins could be replaced by generic functions: this is why I would not provide mixins, should I write a new language from scratch. Of course, in an existing language like Python, one has to follow the common idioms, so I use mixins in a few controlled cases, and I have no problems with that. For instance, one could define an EqualityMixin which defines the special methods __eq__ and __ne__, with __ne__ being the logical negation of __eq__ (Python does not do that by default). That would be a fine usage but I don't do that, I prefer to duplicate two lines of code and to write the __ne__ method explicitly, to avoid complicating my inheritance hierarchy. One should should decide when to use a mixin or not on a case by case basis, with a bias for the not.

Talk Back!

Have an opinion? Readers have already posted 10 comments about this weblog entry. Why not add yours?

RSS Feed

If you'd like to be notified whenever Michele Simionato adds a new entry to his weblog, subscribe to his RSS feed.

About the Blogger

Michele Simionato started his career as a Theoretical Physicist, working in Italy, France and the U.S. He turned to programming in 2003; since then he has been working professionally as a Python developer and now he lives in Milan, Italy. Michele is well known in the Python community for his posts in the newsgroup(s), his articles and his Open Source libraries and recipes. His interests include object oriented programming, functional programming, and in general programming metodologies that enable us to manage the complexity of modern software developement.

This weblog entry is Copyright © 2009 Michele Simionato. All rights reserved.

Sponsored Links



Google
  Web Artima.com   

Copyright © 1996-2019 Artima, Inc. All Rights Reserved. - Privacy Policy - Terms of Use