AddOns 0.7

Dynamically extend other objects with AddOns (formerly ObjectRoles)

(NEW in version 0.6: the``Registry`` base class, and the ClassAddOn.for_frame() classmethod.)

In any sufficiently-sized application or framework, it’s common to end up lumping a lot of different concerns into the same class. For example, you may have business logic, persistence code, and UI all jammed into a single class. Attribute and method names for all sorts of different operations get shoved into a single namespace – even when using mixin classes.

Separating concerns into different objects, however, makes it easier to write reusable and separately-testable components. The AddOns package (peak.util.addons) lets you manage concerns using AddOn classes.

AddOn classes are like dynamic mixins, but with their own private attribute and method namespaces. A concern implemented using add-ons can be added at runtime to any object that either has a writable __dict__ attribute, or is weak-referenceable.

AddOn classes are also like adapters, but rather than creating a new instance each time you ask for one, an existing instance is returned if possible. In this way, add-ons can keep track of ongoing state. For example, a Persistence add-on might keep track of whether its subject has been saved to disk yet:

>>> from peak.util.addons import AddOn

>>> class Persistence(AddOn):
...     saved = True
...     def changed(self):
...         self.saved = False
...     def save_if_needed(self):
...         if not self.saved:
...             print "saving"
...             self.saved = True

>>> class Thing: pass
>>> aThing = Thing()

>>> Persistence(aThing).saved
True
>>> Persistence(aThing).changed()
>>> Persistence(aThing).saved
False
>>> Persistence(aThing).save_if_needed()
saving
>>> Persistence(aThing).save_if_needed() # no action taken

This makes it easy for us to, for example, write a loop that saves a bunch of objects, because we don’t need to concern ourselves with initializing the state of the persistence add-on. A class doesn’t need to inherit from a special base in order to be able to have this state tracked, and it doesn’t need to know how to initialize it, either.

Of course, in the case of persistence, a class does need to know when to call the persistence methods, to indicate changedness and to request saving. However, a library providing such an add-on can also provide decorators and other tools to make this easier, while still remaining largely independent of the objects involved.

Indeed, the AddOns library was actually created to make it easier to implement functionality using function or method decorators. For example, one can create a @synchronized decorator that safely locks an object – see the example below under Threading Concerns.

In summary, the AddOns library provides you with a basic form of AOP, that lets you attach (or “introduce”, in AspectJ terminology) additional attributes and methods to an object, using a private namespace. (If you also want to do AspectJ-style “advice”, the PEAK-Rules package can be used to do “before”, “after”, and “around” advice in combination with add-ons.)

Table of Contents

• Basic API

• AddOn Keys and Instances

• Role Storage and Garbage Collection

• Class Add-Ons

  • Delayed Initialization

  • Keys, Decoration, and for_enclosing_class()

  • Class Registries (NEW in version 0.6)

• Threading Concerns

• Mailing List

Basic API

If you need to, you can query for the existence of an add-on:

>>> Persistence.exists_for(aThing)
True

And by default, it won’t exist:

>>> anotherThing = Thing()
>>> Persistence.exists_for(anotherThing)
False

Until you refer to it directly, e.g.:

>>> Persistence(aThing) is Persistence(anotherThing)
False

At which point it will of course exist:

>>> Persistence.exists_for(anotherThing)
True

And maintain its state, linked to its subject:

>>> Persistence(anotherThing) is Persistence(anotherThing)
True

Until/unless you delete it (or its subject is garbage collected):

>>> Persistence.delete_from(anotherThing)
>>> Persistence.exists_for(anotherThing)
False

AddOn Keys and Instances

Add-ons are stored either in their subject’s __dict__, or if it does not have one (or is a type object with a read-only __dict__), they are stored in a special dictionary linked to the subject via a weak reference.

By default, the dictionary key is the add-on class, so there is exactly one add-on instance per subject:

>>> aThing.__dict__
{<class 'Persistence'>: <Persistence object at...>}

But in some cases, you may wish to have more than one instance of a given add-on class for a subject. (For example, PEAK-Rules uses add-ons to represent indexes on different expressions contained within rules.) For this purpose, you can redefine your AddOn’s __init__ method to accept additional arguments besides its subject. The additional arguments become part of the key that instances are stored under, such that more than one add-on instance can exist for a given object:

>>> class Index(AddOn, dict):
...     def __init__(self, subject, expression):
...         self.expression = expression

>>> something = Thing()
>>> Index(something, "x>y")["a"] = "b"
>>> dir(something)
['__doc__', '__module__', (<class 'Index'>, 'x>y')]

>>> "a" in Index(something, "z<22")
False

>>> Index(something, "x>y")
{'a': 'b'}

>>> Index(something, "x>y").expression
'x>y'

>>> dir(something)
['__doc__', '__module__', (<class 'Index'>, 'x>y'), (<class 'Index'>, 'z<22')]

>>> Index.exists_for(something, 'x>y')
True

>>> Index.exists_for(anotherThing, 'q==42')
False

By default, an add-on class’ key is either the class by itself, or a tuple containing the class, followed by any arguments that appeared in the constructor call after the add-on’s subject. However, you can redefine the addon_key() classmethod in your subclass, and change it to do something different. For example, you could make different add-on classes generate overlapping keys, or you could use attributes of the arguments to generate the key. You could even generate a string key, to cause the add-on to be attached as an attribute!:

>>> class Leech(AddOn):
...     def addon_key(cls):
...         return "__leech__"
...     addon_key = classmethod(addon_key)

>>> something = Thing()

>>> Leech(something) is something.__leech__
True

The addon_key method only receives the arguments that appear after the subject in the constructor call. So, in the case above, it receives no arguments. Had we called it with additional arguments, we’d have gotten an error:

>>> Leech(something, 42)
Traceback (most recent call last):
  ...
TypeError: addon_key() takes exactly 1 argument (2 given)

Naturally, your addon_key() and __init__() (and/or __new__()) methods should also agree on how many arguments there can be, and what they mean!

In general, you should include your add-on class (or some add-on class) as part of your key, so as to make collisions with other people’s add-on classes impossible. Keys should also be designed for thread-safety, where applicable. (See the section below on Threading Concerns for more details.)

Role Storage and Garbage Collection

By the way, the approach above of using an string as an add-on key won’t always make the add-on into an attribute of the subject! If an object doesn’t have a __dict__, or that __dict__ isn’t writable (as in the case of type objects), then the add-on is stored in a weakly-keyed dictionary, maintained elsewhere:

>>> class NoDict(object):
...     __slots__ = '__weakref__'

>>> dictless = NoDict()

>>> Leech(dictless)
<Leech object at ...>

>>> dictless.__leech__
Traceback (most recent call last):
  ...
AttributeError: 'NoDict' object has no attribute '__leech__'

Of course, if an object doesn’t have a dictionary and isn’t weak-referenceable, there’s simply no way to store an add-on for it:

>>> ob = object()
>>> Leech(ob)
Traceback (most recent call last):
  ...
TypeError: cannot create weak reference to 'object' object

However, there is an addons_for() function in the peak.util.addons module that you can extend using PEAK-Rules advice. Once you add a method to support a type that otherwise can’t be used with add-ons, you should be able to use any and all kinds of add-on objects with that type. (Assuming, of course, that you can implement a suitable storage mechanism!)

Finally, a few words regarding garbage collection. If you don’t want to create a reference cycle, don’t store a reference to your subject in your add-on. Even though the __init__ and __new__ messages get the subject passed in, you are not under any obligation to store the subject, and often won’t need to. Usually, the code that is accessing the add-on knows what subject is in use, and can pass the subject to the add-on’s methods if needed. It’s rare that the add-on really needs to keep a reference to the subject past the __new__() and __init__() calls.

Add-on instances will usually be garbage collected at the same time as their subject, unless there is some other reference to them. If they keep a reference to their subject, their garbage collection may be delayed until Python’s cycle collector is run. But if they don’t keep a reference, they will usually be deleted as soon as the subject is:

>>> def deleting(r):
...     print "deleting", r

>>> from weakref import ref

>>> r = ref(Leech(something), deleting)
>>> del something
deleting <weakref at ...; dead>

(Add-ons that are stored outside the instance dictionary of their subject, however, may take slightly longer, as Python processes weak reference callbacks.)

It is also not recommended that you have __del__ methods on your add-on objects, especially if you keep a reference to your subject. In such a case, garbage collection may become impossible, and both the add-on and its subject would “leak” (i.e., take up memory forever without being recoverable).

Class Add-Ons

Sometimes, it’s useful to attach add-ons to classes instead of instances. You could use normal AddOn classes, of course, as they work just fine with both classic classes and new-style types – even built-ins:

>>> Persistence.exists_for(int)
False

>>> Persistence(int) is Persistence(int)
True

>>> Persistence.exists_for(int)
True

>>> class X: pass

>>> Persistence.exists_for(X)
False

>>> Persistence(X) is Persistence(X)
True

>>> Persistence.exists_for(X)
True

But, sometimes you have add-ons that are specifically intended for adding metadata to classes – perhaps by way of class or method decorators. In such a case, you need a way to access the add-on before its subject even exists!

The ClassAddOn base class provides a mechanism for this. It adds an extra classmethod, for_enclosing_class(), that you can use to access the add-on for the class that is currently being defined in the scope that invoked the caller. For example, suppose we want to have a method decorator that adds the method to some class-level registry:

>>> from peak.util.addons import ClassAddOn

>>> class SpecialMethodRegistry(ClassAddOn):
...     def __init__(self, subject):
...         self.special_methods = {}
...         super(SpecialMethodRegistry, self).__init__(subject)

>>> def specialmethod(func):
...     smr = SpecialMethodRegistry.for_enclosing_class()
...     smr.special_methods[func.__name__] = func
...     return func

>>> class Demo:
...     def dummy(self, foo):
...         pass
...     dummy = specialmethod(dummy)

>>> SpecialMethodRegistry(Demo).special_methods
{'dummy': <function dummy at ...>}

>>> class Demo2(object):
...     def dummy(self, foo):
...         pass
...     dummy = specialmethod(dummy)

>>> SpecialMethodRegistry(Demo2).special_methods
{'dummy': <function dummy at ...>}

You can of course use the usual add-on API for class add-ons:

>>> SpecialMethodRegistry.exists_for(int)
False

>>> SpecialMethodRegistry(int).special_methods['x'] = 123

>>> SpecialMethodRegistry.exists_for(int)
True

Except that you cannot explicitly delete them, they must be garbage collected naturally:

>>> SpecialMethodRegistry.delete_from(Demo)
Traceback (most recent call last):
  ...
TypeError: ClassAddOns cannot be deleted

Delayed Initialization

When a class add-on is initialized, the class may not exist yet. In this case, None is passed as the first argument to the __new__ and __init__ methods. You must be able to handle this case correctly, if your add-on will be accessed inside a class definition with for_enclosing_class().

You can, however, define a created_for() instance method that will be called as soon as the actual class is available. It is also called by the default __init__ method, if the add-on is initially created for a class that already exists. Either way, the created_for() method should be called at most once for any given add-on instance. For example:

>>> class SpecialMethodRegistry(ClassAddOn):
...     def __init__(self, subject):
...         print "init called for", subject
...         self.special_methods = {}
...         super(SpecialMethodRegistry, self).__init__(subject)
...
...     def created_for(self, cls):
...         print "created for", cls.__name__

>>> class Demo:
...     def dummy(self, foo):
...         pass
...     dummy = specialmethod(dummy)
init called for None
created for Demo

Above, __init__ was called with None since the type didn’t exist yet. However, accessing the add-on for an existing type (that doesn’t have the add- on yet) will call __init__ with the type, and the default implementation of ClassAddOn.__init__ will also call created_for() for us, when it sees the subject is not None:

>>> SpecialMethodRegistry(float)
init called for <type 'float'>
created for float
<SpecialMethodRegistry object at ...>

>>> SpecialMethodRegistry(float)    # created_for doesn't get called again
<SpecialMethodRegistry object at ...>

One of the most useful features of having this created_for() method is that it allows you to set up class-level metadata that involves inherited settings from base classes. In created_for(), you have access to the class’ __bases__ and or __mro__, and you can just ask for an instance of the same add-on for those base classes, then incorporate their data into your own instance as appropriate. You are guaranteed that any such add-ons you access will already be initialized, including having their created_for() method called.

Since this works recursively, and because class add-ons can be attached even to built-in types like object, the work of creating a correct class metadata registry is immensely simplified, compared to having to special case such base classes, check for bases where no metadata was added or defined, etc.

Instead, classes that didn’t define any metadata will just have an add-on instance containing whatever was setup by your add-on’s __init__() method, plus whatever additional data was added by its created_for() method.

Thus, metadata accumulation using class add-ons can actually be simpler than doing the same things with metaclasses, since metaclasses can’t be retroactively added to existing classes. Of course, class add-ons can’t entirely replace metaclasses or base class mixins, but for the things they can do, they are much easier to implement correctly.

Keys, Decoration, and for_enclosing_class()

Class add-ons can have add-on keys, just like regular add-ons, and they’re implemented in the same way. And, you can pass the extra arguments as positional arguments to for_enclosing_class(). For example:

>>> class Index(ClassAddOn):
...     def __init__(self, subject, expr):
...         self.expr = expr
...         self.funcs = []
...         super(Index, self).__init__(subject)

>>> def indexedmethod(expr):
...     def decorate(func):
...         Index.for_enclosing_class(expr).funcs.append(func)
...         return func
...     return decorate

>>> class Demo:
...     def dummy(self, foo):
...         pass
...     dummy = indexedmethod("x*y")(dummy)

>>> Index(Demo, "x*y").funcs
[<function dummy at ...>]

>>> Index(Demo, "y+z").funcs
[]

Note, by the way, that you do not need to use a function decorator to add metadata to a class. You just need to be calling for_enclosing_class() in a function called directly from the class body:

>>> def special_methods(**kw):
...     smr = SpecialMethodRegistry.for_enclosing_class()
...     smr.special_methods.update(kw)

>>> class Demo:
...     special_methods(x=23, y=55)
init called for None
created for Demo

>>> SpecialMethodRegistry(Demo).special_methods
{'y': 55, 'x': 23}

By default, the for_enclosing_class() method assumes is it being called by a function that is being called directly from the class suite, such as a method decorator, or a standalone function call as shown above. But if you make a call from somewhere else, such as outside a class statement, you will get an error:

>>> special_methods(z=42)
Traceback (most recent call last):
  ...
SyntaxError: Class decorators may only be used inside a class statement

Similarly, if you have a function that calls for_enclosing_class(), but then you call that function from another function, it will still fail:

>>> def sm(**kw):
...     special_methods(**kw)

>>> class Demo:
...     sm(x=23, y=55)
Traceback (most recent call last):
  ...
SyntaxError: Class decorators may only be used inside a class statement

This is because for_enclosing_class() assumes the class is being defined two stack levels above its frame. You can change this assumption, however, by using the level keyword argument:

>>> def special_methods(level=2, **kw):
...     smr = SpecialMethodRegistry.for_enclosing_class(level=level)
...     smr.special_methods.update(kw)

>>> def sm(**kw):
...     special_methods(level=3, **kw)

>>> class Demo:
...     sm(x=23)
...     special_methods(y=55)
init called for None
created for Demo

>>> SpecialMethodRegistry(Demo).special_methods
{'y': 55, 'x': 23}

Alternately, you can pass a specific Python frame object via the frame keyword argument to for_enclosing_class(), or use the for_frame() classmethod instead. for_frame() takes a Python stack frame, followed by any extra positional arguments needed to create the key.

Class Registries (NEW in version 0.6)

For many of common class add-on use cases, you just want a dict-like object with “inheritance” for the values in base classes. The Registry base class provides this behavior, by subclassing ClassAddOn and the Python dict builtin type, to create a class add-on that’s also a dictionary. It then overrides the created_for() method to automatically populate itself with any inherited values from base classes.

Let’s define a MethodGoodness registry that will store a “goodness” rating for methods:

>>> from peak.util.addons import Registry

>>> class MethodGoodness(Registry):
...     """Dictionary of method goodness"""

>>> def goodness(value):
...     def decorate(func):
...         MethodGoodness.for_enclosing_class()[func.__name__]=value
...         return func
...     return decorate

>>> class Demo(object):
...     def aMethod(self, foo):
...         pass
...     aMethod = goodness(17)(aMethod)
...     def another_method(whinge, spam):
...         woohoo
...     another_method = goodness(-99)(another_method)

>>> MethodGoodness(Demo)
{'aMethod': 17, 'another_method': -99}

So far, so good. Let’s see what happens with a subclass:

>>> class Demo2(Demo):
...     def another_method(self, fixed):
...         pass
...     another_method = goodness(42)(another_method)

>>> MethodGoodness(Demo2)
{'another_method': 42, 'aMethod': 17}

Values set in base class registries are automatically added to the current class’ registry of the same type and key, if the current class doesn’t have an entry defined. Python’s new-style method resolution order is used to determine the precedence of inherited attributes. (For classic classes, a temporary new-style class is created that inherits from the classic class, in order to determine the resolution order, then discarded.)

Once the class in question has been created, the registry gets an extra attribute, defined_in_class, which is a dictionary listing the entries that were actually defined in the corresponding class, e.g.:

>>> MethodGoodness(Demo).defined_in_class
{'aMethod': 17, 'another_method': -99}

>>> MethodGoodness(Demo2).defined_in_class
{'another_method': 42}

As you can see, this second dictionary contains only the values registered in that class, and not any inherited values.

Finally, note that Registry objects have one additional method that can be useful to call from a decorator: set(key, value). This method will raise an error if a different value already exists for the given key, and is useful for catching errors in class definitions, e.g.:

>>> def goodness(value):
...     def decorate(func):
...         MethodGoodness.for_enclosing_class().set(func.__name__, value)
...         return func
...     return decorate

>>> class Demo3(object):
...     def aMethod(self, foo):
...         pass
...     aMethod = goodness(17)(aMethod)
...     def aMethod(self, foo):
...         pass
...     aMethod = goodness(27)(aMethod)
Traceback (most recent call last):
  ...
ValueError: MethodGoodness['aMethod'] already contains 17; can't set to 27

Threading Concerns

Add-on lookup and creation is thread-safe (i.e. race-condition free), so long as the add-on key contains no objects with __hash__ or __equals__ methods involve any Python code (as opposed to being pure C code that doesn’t call any Python code). So, unkeyed add-ons, or add-ons whose keys consist only of instances of built-in types (recursively, in the case of tuples) or types that inherit their __hash__ and __equals__ methods from built-in types, can be initialized in a thread-safe manner.

This does not mean, however, that two or more add-on instances can’t be created for the same subject at the same time! Code in an add-on class’ __new__ or __init__ methods must not assume that it will in fact be the only add-on instance attached to its subject, if you wish the code to be thread-safe.

This is because the AddOn access machinery allows multiple threads to create an add-on instance at the same time, but only one of those objects will win the race to become “the” add-on instance, and no thread can know in advance whether it will win. Thus, if you wish your AddOn instances to do something to their constructor arguments at initialization time, you must either give up on your add-on being thread-safe, or use some other locking mechanism.

Of course, add-on initialization is only one small part of the overall thread- safety puzzle. Unless your add-on exists only to compute some immutable metadata about its subject, the rest of your add-on’s methods need to be thread-safe also.

One way to do that, is to use a @synchronized decorator, combined with a Locking add-on:

>>> class Locking(AddOn):
...     def __init__(self, subject):
...         from threading import RLock
...         self.lock = RLock()
...     def acquire(self):
...         print "acquiring"
...         self.lock.acquire()
...     def release(self):
...         self.lock.release()
...         print "released"

>>> def synchronized(func):
...     def wrapper(self, *__args,**__kw):
...         Locking(self).acquire()
...         try:
...             func(self, *__args,**__kw)
...         finally:
...             Locking(self).release()
...
...     from peak.util.decorators import rewrap
...     return rewrap(func, wrapper)

>>> class AnotherThing:
...     def ping(self):
...         print "ping"
...     ping = synchronized(ping)

>>> AnotherThing().ping()
acquiring
ping
released

If the Locking() add-on constructor were not thread-safe, this decorator would not be able to do its job correctly, because two threads accessing an object that didn’t have the add-on yet, could end up locking two different locks, and proceeding to run the supposedly-“synchronized” method at the same time!

(In general, thread-safety is harder than it looks. But at least you don’t have to worry about this one tiny part of correctly implementing it.)

Of course, synchronized methods will be slower than normal methods, which is why AddOns doesn’t do anything besides that one small part of the thread-safety puzzle, to avoid penalizing non-threaded code. As the PEAK motto says, STASCTAP! (Simple Things Are Simple, Complex Things Are Possible.)

Mailing List

Questions, discussion, and bug reports for this software should be directed to the PEAK mailing list; see http://www.eby-sarna.com/mailman/listinfo/PEAK/ for details.