skip to navigation
skip to content

Not Logged In

strait 0.5.1

Simple Traits for Python



<html xmlns="http://www.w3.org/1999/xhtml" xml:lang="en" lang="en">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8"/>
<meta name="generator" content="Docutils 0.4.1: http://docutils.sourceforge.net/"/>
<title>A simple implementation of traits for Python</title>
<meta name="author" content="Michele Simionato"/>
<meta name="date" content="2009-01-10"/>
<style type="text/css">

.highlight { background: #f8f8f8; }
.highlight .c { color: #408080; font-style: italic } /* Comment */
.highlight .err { border: 1px solid #FF0000 } /* Error */
.highlight .k { color: #008000; font-weight: bold } /* Keyword */
.highlight .o { color: #666666 } /* Operator */
.highlight .cm { color: #408080; font-style: italic } /* Comment.Multiline */
.highlight .cp { color: #BC7A00 } /* Comment.Preproc */
.highlight .c1 { color: #408080; font-style: italic } /* Comment.Single */
.highlight .cs { color: #408080; font-style: italic } /* Comment.Special */
.highlight .gd { color: #A00000 } /* Generic.Deleted */
.highlight .ge { font-style: italic } /* Generic.Emph */
.highlight .gr { color: #FF0000 } /* Generic.Error */
.highlight .gh { color: #000080; font-weight: bold } /* Generic.Heading */
.highlight .gi { color: #00A000 } /* Generic.Inserted */
.highlight .go { color: #808080 } /* Generic.Output */
.highlight .gp { color: #000080; font-weight: bold } /* Generic.Prompt */
.highlight .gs { font-weight: bold } /* Generic.Strong */
.highlight .gu { color: #800080; font-weight: bold } /* Generic.Subheading */
.highlight .gt { color: #0040D0 } /* Generic.Traceback */
.highlight .kc { color: #008000; font-weight: bold } /* Keyword.Constant */
.highlight .kd { color: #008000; font-weight: bold } /* Keyword.Declaration */
.highlight .kp { color: #008000 } /* Keyword.Pseudo */
.highlight .kr { color: #008000; font-weight: bold } /* Keyword.Reserved */
.highlight .kt { color: #008000; font-weight: bold } /* Keyword.Type */
.highlight .m { color: #666666 } /* Literal.Number */
.highlight .s { color: #BA2121 } /* Literal.String */
.highlight .na { color: #7D9029 } /* Name.Attribute */
.highlight .nb { color: #008000 } /* Name.Builtin */
.highlight .nc { color: #0000FF; font-weight: bold } /* Name.Class */
.highlight .no { color: #880000 } /* Name.Constant */
.highlight .nd { color: #AA22FF } /* Name.Decorator */
.highlight .ni { color: #999999; font-weight: bold } /* Name.Entity */
.highlight .ne { color: #D2413A; font-weight: bold } /* Name.Exception */
.highlight .nf { color: #0000FF } /* Name.Function */
.highlight .nl { color: #A0A000 } /* Name.Label */
.highlight .nn { color: #0000FF; font-weight: bold } /* Name.Namespace */
.highlight .nt { color: #008000; font-weight: bold } /* Name.Tag */
.highlight .nv { color: #19177C } /* Name.Variable */
.highlight .ow { color: #AA22FF; font-weight: bold } /* Operator.Word */
.highlight .w { color: #bbbbbb } /* Text.Whitespace */
.highlight .mf { color: #666666 } /* Literal.Number.Float */
.highlight .mh { color: #666666 } /* Literal.Number.Hex */
.highlight .mi { color: #666666 } /* Literal.Number.Integer */
.highlight .mo { color: #666666 } /* Literal.Number.Oct */
.highlight .sb { color: #BA2121 } /* Literal.String.Backtick */
.highlight .sc { color: #BA2121 } /* Literal.String.Char */
.highlight .sd { color: #BA2121; font-style: italic } /* Literal.String.Doc */
.highlight .s2 { color: #BA2121 } /* Literal.String.Double */
.highlight .se { color: #BB6622; font-weight: bold } /* Literal.String.Escape */
.highlight .sh { color: #BA2121 } /* Literal.String.Heredoc */
.highlight .si { color: #BB6688; font-weight: bold } /* Literal.String.Interpol */
.highlight .sx { color: #008000 } /* Literal.String.Other */
.highlight .sr { color: #BB6688 } /* Literal.String.Regex */
.highlight .s1 { color: #BA2121 } /* Literal.String.Single */
.highlight .ss { color: #19177C } /* Literal.String.Symbol */
.highlight .bp { color: #008000 } /* Name.Builtin.Pseudo */
.highlight .vc { color: #19177C } /* Name.Variable.Class */
.highlight .vg { color: #19177C } /* Name.Variable.Global */
.highlight .vi { color: #19177C } /* Name.Variable.Instance */
.highlight .il { color: #666666 } /* Literal.Number.Integer.Long */

</style>
</head>
<body>

A simple implementation of traits for Python

















Author:Michele Simionato
Date:2009-01-10
Version:0.5.1
Download:http://pypi.python.org/pypi/strait
Licence:BSD


Abstract


I provide a simple implementation of
traits as units of composable behavior for Python. I
argue that traits are better than multiple inheritance.
Implementing frameworks based on traits is left as an exercise
for the reader.




Motivation


Multiple inheritance is a hotly debated topic.
The supporters of multiple inheritance
claim that it makes code shorter and easier
to read, whereas the opposers claim that is makes
code more coupled and more difficult to understand. I have
spent some time in the past facing the intricacies of multiple
inheritance in Python
and I was one of its supporters once; however,
since then I have worked with frameworks making
large use of multiple inheritance (I mean Zope 2) and nowadays I am in
the number of the people who oppose it. Therefore I
am interested in alternatives.


In recent years, the approach
of traits has gained some traction in a few circles and I have
decided to write a library to implement traits in Python, for
experimentation purposes. The library is meant for framework builders,
people who are thinking about writing a framework based on multiple
inheritance - typically via the common mixin approach - but
are not convinced that this is the best solution and would like to try
an alternative. This library is also for authors of mixin-bases frameworks
which are unsatisfied and would like to convert their
framework to traits.


Are traits a better solution than multiple inheritance and mixins? In
theory I think so, otherwise I would not have written this library, but
in practice (as always) things may be different. It may well be
that using traits or using mixins does not make a big
difference in practice
and that the change of paradigm is not worth the effort; or the
opposite may be true. The only way to know is to try, to build
software based on traits and to see how it scale in the large.
In the small, more or less any approach works fine: it is only
by programming in the large that you can see the differences.


This is
the reason why I am releasing this library with a liberal licence, so
that people can try it out and see how it works. The library is meant
to play well (when possible) with pre-existing frameworks.
As an example, I will show
here how you could rewrite Tkinter classes to use traits instead of mixins. Of
course, I am not advocating rewriting Tkinter: it would be silly
and pointless; but it may have sense (or
not) to rewrite your own framework using traits, perhaps a framework
which is used in house but has not been released yet.


I am not the only one to have
implemented traits for Python; after finishing my implementation
I made a little research and discovered a few implementations. Then
I have also discovered the Enthought Traits framework, which however
seems to use the name to intend something completely
different (i.e. a sort of type checking). My implementation has no
dependencies, is short and I am committed
to keep it short even in the future, according to
the principle of less is more.


There is also an hidden agenda behind this module: to popularize some
advanced features of the Python object model which are little
known. The strait module is actually a tribute to the
metaprogramming capabilities of Python: such features are usually
associated to languages with a strong academic tradition - Smalltalk,
Scheme, Lisp - but actually the Python object model is no less
powerful. For instance, changing the object system from a multiple
inheritance one to a trait-based one,
can be done within the fundamental object system. The reason is that
the features that Guido used to implement the object system (special
method hooks, descriptors, metaclasses) are there, available to the
end user to build her own object system.


Such features are usually little used in the Python community, for
many good reasons: most people feel that the object system is good
enough and that there is no reason to change it; moreover there is
a strong opposition to change the language, because Python programmers
believe in uniformity and in using common idioms; finally, it is
difficult for an application programmer to find a domain where these
features are useful. An exception is the domain of the Object Relation
Mappers, whereas the Python language is often stretched to mimic the
SQL language, a famous example of this tendency being SQLAlchemy).
Still, I have never seen a perversion of the object model as big as
the one implemented in the strait module, so I wanted to be the
first one to perform that kind of abuse ;)




What are traits?


The word traits has many meanings; I will refer to it in the sense
of the paper Traits - Composable Units of Behavior which implements
them in Squeak/Smalltalk. The paper appeared in 2003, but most of the
ideas underlying traits have been floating around for at least 30
years. There is also a trait implementation for PLT Scheme which is
somewhat close in spirit (if not in practice) to what I am advocating here.
The library you are reading about is by no means intended as a porting
of the Smalltalk library: I am just stealing some of the ideas from
that paper to implement a Pythonic alternative to mixins which, for
lack of a better name, I have decided to call traits. I feel no
obligation whatsoever to be consistent with the Smalltalk library. In
doing so, I am following a long tradition, since a lot of languages
use the name traits to mean something completely different from the
Smalltalk meaning. For instance the languages Fortress and Scala use
the name trait but with a different meaning (Scala traits are very
close to multiple inheritance).
For me a trait is a bunch of methods and attributes with the following
properties:



  1. the methods/attributes in a trait belong logically together;

  2. if a trait enhances a class, then all subclasses are enhanced too;

  3. if a trait has methods in common with the class, then the
    methods defined in the class have the precedence;

  4. the trait order is not important, i.e. enhancing a class
    first with trait T1 and then with trait T2 or viceversa is the same;

  5. if traits T1 and T2 have names in common, enhancing a class both
    with T1 and T2 raises an error;

  6. if a trait has methods in common with the base class, then the
    trait methods have the precedence;

  7. a class can be seen both as a composition of traits and as an homogeneous
    entity.


Properties from 4 to 7 are the distinguishing properties of traits
with respect to multiple inheritance and mixins. In particular,
because of 4 and 5, all the complications with the Method Resolution
Order disappear and the overriding is never implicit. Property 6 is
mostly unusual: typically in Python the base class has the precedence
over mixin classes. Property 7 should be intended in the sense that a
trait implementation must provide introspection facilities to make
seemless the transition between classes viewed as atomic entities and
as composed entities.




A hands-on example


Let me begin by showing how you could rewrite a
Tkinter class to use traits instead of mixins. Consider the
Tkinter.Widget class, which is derived by the base class
BaseWidget and the mixin classes
Tkinter.Grid, Tkinter.Pack and Tkinter.Place: I want to
rewrite it by using traits. The strait module
provides a factory function named include that does the job.
It is enough to replace the multiple inheritance syntax:


class Widget(BaseWidget, Grid, Pack, Place):
pass

with the following syntax:


class Widget(BaseWidget):
__metaclass__ = include(Pack, Place, Grid)

I said that the conversion from mixins to traits was easy: but actually
I lied since if you try to execute the code I just wrote you will
get an OverridingError:



>>> from Tkinter import *
>>> class Widget(BaseWidget):
... __metaclass__ = include(Pack, Place, Grid)
Traceback (most recent call last):
...
OverridingError: Pack overrides names in Place: {info, config, configure,
slaves, forget}


The reason for the error is clear: both Pack and Place provide
methods called {info, config, configure, slaves, forget}
and the traits implementation cannot figure out
which ones to use. This is a feature, since it forces you to be
explicit. In this case, if we want to be consistent with
multiple inheritance rules, we want the methods coming from
the first class (i.e. Pack) to take precedence. That can be
implemented by including directly those methods in the class namespace
and relying on rule 3:



class TOSWidget(BaseWidget):
__metaclass__ = include(Pack, Place, Grid)
info = Pack.info.im_func
config = Pack.config.im_func
configure = Pack.configure.im_func
slaves = Pack.slaves.im_func
forget = Pack.forget.im_func
propagate = Pack.propagate.im_func



Notice that we had to specify the propagate method too, since
it is a common method between Pack and Grid.


You can check that the TOSWidget class works, for instance by defining a
label widget as follows (remember that TOSWidget inherits its signature
from BaseWidget):



>>> label = TOSWidget(master=None, widgetName='label',
... cnf=dict(text="hello"))


You may visualize the widget by calling the .pack method:



>>> label.pack()


This should open a small window with the message "hello" inside it.




A few caveats and warnings


First of all, let me notice that, in spite of apparency, include
does not return a metaclass. Insted, it returns a class factory
function with signature name, bases, dic:


>>> print include(Pack, Place, Grid)
<function include_Pack_Place_Grid at 0x...>

This function will create the class by using a suitable
metaclass:


>>> type(TOSWidget)
<class 'strait.MetaTOS'>

In simple cases the metaclass will be MetaTOS, the main class
of the trait object system, but in general it can be a different
one not inheriting from MetaTOS. The exact rules followed by
include to determine the right class will be discussed later.


Here I want to remark that according to rule 6 traits take the precedence
over the base class attributes. Consider the following example:



>>> class Base(object):
... a = 1

>>> class ATrait(object):
... a = 2

>>> class Class(Base):
... __metaclass__ = include(ATrait)

>>> Class.a
2


In regular multiple inheritance you would do the same by including
ATrait before Base, i.e.



>>> type('Class', (ATrait, Base), {}).a
2


You should take care to not mix-up the order, otherwise you will get a
different result:



>>> type('Class', (Base, ATrait), {}).a
1


Therefore replacing mixin classes with traits can break your code if
you rely on the order. Be careful!




The Trait Object System


The goal of the strait module it to modify the standard
Python object model, turning it into a Trait Object System (TOS for short):
TOS classes behave differently from regular
classes. In particular TOS classes do not support multiple inheritance.
If you try to multiple inherit from a TOS
class and another class you will get a TypeError:



>>> class M:
... "An empty class"
...
>>> class Widget2(TOSWidget, M):
... pass
...
Traceback (most recent call last):
...
TypeError: Multiple inheritance of bases (<class '__main__.TOSWidget'>, <class
__main__.M at 0x...>) is forbidden for TOS classes


This behavior is intentional: with this restriction you can simulate
an ideal world in which Python did not support multiple
inheritance. Suppose you want to claim that supporting multiple
inheritance was a mistake and that Python would have been better off
without it (which is the position I tend to have nowadays): how can
you prove that claim? Simply by writing code that does not use
multiple inheritance and it is clearer and more mantainable that code
using multiple inheritance.


I am releasing this trait implementation hoping you will help me to
prove (or possibly disprove) the point. You may see traits as a
restricted form of multiple inheritance without name clashes,
without the complications of the method resolution, and with a
limited cooperation between methods.
Moreover the present implementation is slightly less dynamic
than usual inheritance.


A nice property of inheritance is that if you have a class C inheriting
from class M and you change a method in M at runtime, after
C has been created and instantiated, automagically all instances
of C gets the new version of the method, which is pretty useful
for debugging purposes. This feature is lost in the trait implementation
provided here. Actually, in a previous version, my trait implementation
was fully dynamic and if you changed the mixin the instances would be
changed too. However, I never used that feature in practice, and
it was complicating the implementation and slowing doing the
attribute access, so I removed it.


I think these are acceptable restrictions since they give back
in return many advantages in terms of simplicity: for instance,
super becomes trivial, since each class has a single superclass,
whereas we all know that the current super in Python is very far
from trivial.




The magic of include


Since the fundamental properties of TOS classes must be preserved under
inheritance (i.e. the son of a TOS class must be a TOS class)
the implementation necessarily requires metaclasses. As of now,
the only fundamental property of a TOS class is that multiple
inheritance is forbidden, so usually (but not always) TOS
classes are instances of the metaclass MetaTOS
which implements a single inheritance check.
If you build your TOS hierarchy starting from pre-existing classes,
you should be aware of how include determines the metaclass:
if your base class was an old-style
class or a plain new style class (i.e. a direct instance of the
type metaclass), them include will change it to MetaTOS:


>>> type(TOSWidget)
<class 'strait.MetaTOS'>

In general you may need to build your Trait Based Framework
on top of pre-existing classes possessing a nontrivial metaclass, for
instance Zope classes; in that case include is smart
enough to figure out the right metaclass to use. Here is an example:



class AddGreetings(type):
"A metaclass adding a 'greetings' attribute for exemplification purposes"
def __new__(mcl, name, bases, dic):
dic['greetings'] = 'hello!'
return super(AddGreetings, mcl).__new__(mcl, name, bases, dic)




class WidgetWithGreetings(BaseWidget, object):
__metaclass__ = AddGreetings




class PackWidget(WidgetWithGreetings):
__metaclass__ = include(Pack)



include automatically generates the right metaclass as
a subclass of AddGreetings:



>>> print type(PackWidget).__mro__
(<class 'strait._TOSAddGreetings'>, <class '__main__.AddGreetings'>, <type
'type'>, <type 'object'>)


Incidentally, since TOS
classes are guaranteed to be in a straight hierarchy, include is able
to neatly avoid the dreaded metaclass conflict.


The important point is that _TOSAddGreetings provides the same features of
MetaTOS, even if it is not a subclass of it; on the
other hand, _TOSMetaAddGreetings is a subclass of AddGreetings
which calls AddGreetings.__new__, so the features provided by
AddGreetings are not lost either; in this example you may check
that the greetings attribute is correctly set:



>>> PackWidget.greetings
'hello!'


The name of the generated metaclass
is automatically generated from the name of the base
metaclass; moreover, a register of the generated metaclasses
is kept, so that metaclasses are reused if possible.
If you want to understand the details, you are welcome
to give a look at the implementation, which is pretty short
and simple, compared to the general recipe to remove
the metaclass conflict in a true multiple inheritance situation.




Cooperative traits


At first sight, the Trait Object System lacks an important feature of
multiple inheritance as implemented in the ordinary Python object system,
i.e. cooperative methods. Consider for instance the following
classes:



class LogOnInitMI(object):
def __init__(self, *args, **kw):
print 'Initializing %s' % self
super(LogOnInitMI, self).__init__(*args, **kw)




class RegisterOnInitMI(object):
register = []
def __init__(self, *args, **kw):
print 'Registering %s' % self
self.register.append(self)
super(RegisterOnInitMI, self).__init__(*args, **kw)



In multiple inheritance LogOnInitMI can be mixed with other
classes, giving to the children the ability to log on initialization;
the same is true for RegisterOnInitMI, which gives to its children
the ability to populate a registry of instances. The important feature
of the multiple inheritance system is that LogOnInitMI and
RegisterOnInitMI play well together: if you inherits from
both of them, you get both features:



class C_MI(LogOnInitMI, RegisterOnInitMI):
pass



>>> c = C_MI()
Initializing <__main__.C_MI object at 0x...>
Registering <__main__.C_MI object at 0x...>

You cannot get the same behaviour if you use the trait object system
naively:


>>> class C_MI(object):
... __metaclass__ = include(LogOnInitMI, RegisterOnInitMI)
...
Traceback (most recent call last):
...
OverridingError: LogOnInitMI overrides names in RegisterOnInitMI: {__init__}

This is a feature, of course, since the trait object system is designed
to avoid name clashes. However, the situation is worse than that:
even if you try to mixin a single class you will run into trouble


>>> class C_MI(object):
... __metaclass__ = include(LogOnInitMI)

>>> c = C_MI()
Traceback (most recent call last):
...
TypeError: super(type, obj): obj must be an instance or subtype of type

What's happening here? The situation is clear if you notice that the
super call is actually a call of kind super(LogOnInitMI, c)
where c is an instance of C, which is not a
subclass of LogOnInitMI. That explains the
error message, but does not explain how to solve the issue. It seems
that method cooperation using super is impossible for TOS
classes.


Actually this is not the case: single inheritance cooperation
is possible and it is enough as we will show in a
minute. But for the moment let me notice that I do not think
that cooperative methods are necessarily a good idea. They are
fragile and cause all of your classes to be strictly coupled. My usual
advice if that you should not use a design based on method
cooperation if you can avoid it.
Having said that, there are situations (very rare) where you
really want method cooperation. The strait module provide
support for those situations via the __super attribute.


Let me explain how it works. When you mix-in a trait T into a
class C, include adds an attribute _T__super to C,
which is a super object that dispatches to the attributes of the
superclass of C. The important thing to keep in mind is that there
is a well defined superclass, since the trait object system uses
single inheritance only. Since the hierarchy is straight, the
cooperation mechanism is much simpler to understand than in multiple
inheritance. Here is an example. First of all, let me rewrite
LogOnInit and RegisterOnInit to use __super instead of
super:



class LogOnInit(object):
def __init__(self, *args, **kw):
print 'Initializing %s' % self
self.__super.__init__(*args, **kw)




class RegisterOnInit(object):
register = []
def __init__(self, *args, **kw):
print 'Registering %s' % self
self.register.append(self)
self.__super.__init__(*args, **kw)



Now you can include the RegisterOnInit functionality as follows:



class C_Register(object):
__metaclass__ = include(RegisterOnInit)



>>> _ = C_Register()
Registering <__main__.C_Register object at 0x...>

Everything works because include has added the right attribute:


>>> C_Register._RegisterOnInit__super
<super: <class 'C_Register'>, <C_Register object>>

Moreover, you can also include the LogOnInit functionality:



class C_LogAndRegister(C_Register):
__metaclass__ = include(LogOnInit)



>>> _ = C_LogAndRegister()
Initializing <__main__.C_LogAndRegister object at 0x...>
Registering <__main__.C_LogAndRegister object at 0x...>

As you see, the cooperation mechanism works just fine. I will call
cooperative trait a class intended for inclusion in other classes
and making use of the __super trick. A class using the
regular super directly cannot be used as a cooperative trait, since it
must satisfy inheritance constraints, nevertherless it is easy enough to
convert it to use __super. After all, the strait module is
intended for framework writers, so it assumes you can change the
source code of your framework if you want. On the other hand, if
are trying to re-use a mixin class coming from a third party
framework and using super, you will have to rewrite the
parts of it. That is unfortunate, but I cannot perform miracles.


You may see __super as a clever hack to use
super indirectly. Notice that since the hierarchy is straight,
there is room for optimization at the core language
level. The __super trick as implemented in pure Python leverages
on the name mangling mechanism, and follows closely the famous
autosuper recipe, with some improvement. Anyway,
if you have two traits with the same
name, you will run into trouble. To solve this and to have a nicer
syntax, one would need more support from the language, but the
__super trick is good enough for a prototype and
has the serious advantage of working right now for current Python.




Cooperation at the metaclass level


In my experience, the cases where you need method cooperation
in multiple inheritance situations are exceedingly rare,
unless you are a language implementor or a designer of
very advanced frameworks. In such a realm you have a need for
cooperative methods; it is not a pressing need, in the sense that
you can always live without them, but they are a nice feature to have if you
care about elegance and extensibility. For instance, as P. J. Eby
points it out in this thread on python-dev:


A major use case for co-operative super() is in the implementation of
metaclasses. The __init__ and __new__ signatures are fixed, multiple
inheritance is possible, and co-operativeness is a must (as the base
class methods
must be called). I'm hard-pressed to think of a
metaclass constructor or initializer that I've written in the last
half-decade or more where I didn't use super() to make it
co-operative. That, IMO, is a compelling use case even if there were
not a single other example of the need for super.


I have always felt the same. So, even if I have been unhappy with multiple
inheritance for years, I could never dismiss it entirely
because of the concern for this use case. It is only after discovering
cooperative traits that I felt the approach powerful enough
to replace multiple inheritance without losing anything I cared about.


Multiple inheritance at the metaclass level comes out here and
again when you are wearing the language implementor hat. For instance,
if you try to implement an object system based on traits, you will have to do
so at the metaclass level and there method cooperation has its place.
In particular, if you look at the source code of the strait module -
which is around 100 lines, a tribute to the power of Python -
you will see that the MetaTOS metaclass is implemented
as a cooperative trait, so that it can be mixed-in with other metaclasses,
in the case you are interoperating with a framework with a non-trivial
meta object protocol. This is performed internally by include.


Metaclass cooperation is there to make the life of the users
easier. Suppose one of you, users of the strait module, wants to
enhance the include mechanism using another a metaclass coming for
a third party framework and therefore not inheriting from MetaTOS:



class ThirdPartyMeta(type):
def __new__(mcl, name, bases, dic):
print 'Using ThirdPartyMeta to create %s' % name
return super(ThirdPartyMeta, mcl).__new__(mcl, name, bases, dic)



The way to go is simple. First, you should mix-in MetaTOS in the
third party class:



class EnhancedMetaTOS(ThirdPartyMeta):
__metaclass__ = include(MetaTOS)



Then, you can define your own enhanced include as follows:



def enhanced_include(*traits):
return include(MetaTOS=EnhancedMetaTOS, *traits)



In simple cases using directly ThirdPartyMeta may work, but I strongly
recommend to replace the call to super with __super even in
ThirdPartyMeta to make the cooperation robust.




Discussion of some design decisions and future work


The decision of having TOS classes which are not instances of
MetaTOS
required some thought. That was my original idea in version 0.1 of
strait; however in version 0.2 I wanted to see what would happen
if I made all TOS classes instances of MetaTOS.
That implied that if
your original class had a nontrivial metaclass, then the TOS class had
to inherit both from the original metaclass and MetaTOS,
i.e. multiple inheritance and cooperation of methods was required at
the metaclass level.


I did not like it, since I was arguing that
you can do everything without multiple inheritance; moreover using
multiple inheritance at the metaclass level
meant that one had to solve the metaclass conflict in a general
way. I did so, by using my own cookbook recipe, and all my tests
passed.


Neverthess, at the end, in version 0.3 I decided to go back to the
original design. The metaclass conflict recipe is too complex, and I
see it as a code smell - if the implementation is hard to explain,
it's a bad idea
- just another indication that multiple inheritance
is bad. In the original design it is possible to add the features of
MetaTOS to the original metaclass by subclassing it with single
inheritance and thus avoiding the conflict.


The price to pay is that now a TOS class is no more an instance of
MetaTOS, but this is a non-issue: the important
thing is that TOS classes perform the dispatch on their traits as
MetaTOS would dictate. Moreover, starting from
Python 2.6, thanks to Abstract Base Classes, you may satisfy the
isinstance(obj, cls) check even if obj is not an instance of
cls, by registering a suitable base class (similarly for
issubclass). In our situation, that means that it is enough to
register MetaTOS as base class of the original
metaclass.


Version 0.4 was much more complex that the current version (still
short, it was under 300 lines of pure Python), since it had the more
ambitious goal of solving the namespace pollution problem. I have
discussed the issue elsewhere: if you keep injecting methods into a
class (both directly or via inheritance) you may end up having
hundreds of methods flattened at the same level.


A picture is worth a
thousand words, so have a look at the PloneSite hierarchy if you
want to understand the horror I wanted to avoid with traits (the
picture shows the number of nonspecial attributes defined per class in
square brackets): in the Plone Site hierarchy there are 38 classes, 88
overridden names, 42 special names, 648 non-special attributes and
methods. It is a nighmare.


Originally I wanted to prevent this kind
of abuse, but that made my implementation more complex, whereas
my main goal was to keep the implementation simple. As a consequence
this version assume the prosaic attitude that you cannot stop
programmers from bad design anyway, so if they want to go the Zope
way they can.


In previous versions I did provide some syntactic sugar for include
so that it was possible to write something like the following
(using a trick discussed here):


class C(Base):
include(Trait1, Trait2)

In version 0.5 I decided to remove this feature. Now the plumbing
(i.e. the __metaclass__ hook) is exposed to the user, some magic
has been removed and it is easier for the user to write her own
include factory if she wants to.


Where to go from here? For the moment, I have no clear idea about the
future. The Smalltalk implementation of traits provides method
renaming out of the box. The Python implementation has no facilities
in this sense. In the future I may decide to give some support for
renaming, or I may not. At the present you can just rename your
methods by hand. Also, in the future I may decide to add some kind of
adaptation mechanism or I may not: after all the primary goal of this
implementation is semplicity and I don't want to clutter it with too
many features.


I am very open to feedback and criticism: I am releasing this module
with the hope that it will be used in real life situations to gather
experience with the traits concept. Clearly I am not proposing that
Python should remove multiple inheritance in favor of traits:
considerations of backward compatibily would kill the proposal right
from the start. I am just looking for a few adventurous volunteers
wanting to experiment with traits; if the experiment goes well, and
people start using (multiple) inheritance less than they do now, I
will be happy.




Trivia


strait officially stands for Simple Trait object system, however
the name is also a pun on the world "straight", since the difference
between multiple inheritance hierarchies and TOS hierarchies is that
TOS hierarchies are straight. Moreover, nobody will stop you from
thinking that the s also stands for Simionato ;)






View document source.
Generated on: 2009-01-10 10:52 UTC.
Generated by Docutils from reStructuredText source.


</body>
</html>


 
File Type Py Version Uploaded on Size
strait-0.5.1.tar.gz (md5) Source 2009-01-05 101KB
  • Downloads (All Versions):
  • 5 downloads in the last day
  • 39 downloads in the last week
  • 138 downloads in the last month