Charming Python #b17: The Python Enterprise Application Kit.

A first look at protocols in Python


David Mertz, Ph.D.
Your Obedient Writer, Gnosis Software, Inc.
March, 2004

PEAK is a Python framework for rapidly developing and reusing application components. While Python itself is already a very-high-level language, PEAK provides even higher abstractions, largely through clever use of metaclasses and other advanced Python techniques. In many ways PEAK does for Python what J2EE does for Java. Part of the formalization in PEAK comes in the explicit specification of protocols, specifically in the separately available package PyProtocols.

The Python Enterprise Application Kit

If you felt like trying to get a handle on metaclasses, or wrestling with asynchronous programming in Twisted, or thinking about object oriented programming with multiple dispatch, made your brain melt, then you ain't seen nothing yet! PEAK combines some elements of all of these into a component programming framework. PEAK is still rough around the edges. Like Twisted, PEAK's documentation, though extensive, is difficult to penetrate. But for all that, there is something enormously interesting about this project, led by Python guru Phillip J. Eby; and, I think, some opportunities for extremely productive and ultra-high-level application development.

The PEAK package is composed a number of subpackages for different purposes. Some significant subpackages are peak.api, peak.binding, peak.config, peak.naming, and peak.storage. Most of those names are fairly self-explanatory. The subpackage peak.binding is for flexibly connecting components; peak.config lets you store "lazily immutable" data, which are involved in declarative application programming; peak.naming lets you create globally unique identifiers for (networked) resources; peak.storage is pretty much exactly what the name suggests, it lets you manage databases and persistence.

For this article, however, what we will be interested in is peak.api. In particular, the package PyProtocols which is available separately provides an infrastructure for other PEAK subpackages. A version of the PyProtocols package is included within peak.api.protocols. For now though, I am interested in looking at the separate protocols package. In later installments, I will return to other parts of PEAK.

What Is A Protocol?

A protocol, abstractly, is simply a set of behaviors that an object agrees to conform to. Strongly-typed programming languages--including Python--have a collection of basic types, each of which has such a collection of guaranteed behaviors: Integers know how to multiply themselves together; lists know how to iterate over their contents; dictionaries know how to find a value by key; files know how to read and write bytes; etc. The collection of behaviors you can expect out of built-in types constitutes a protocol they implement. An object that codifies an protocol is known as an interface.

For standard types, it is not too difficult to explicitly list all the behaviors implemented (though these change a bit between Python version; or certainly between different programming languages). But at the boundaries--for objects that belong to custom classes--it is difficult to state definitively what constitutes "dictionary-like" or "file-like" behavior. Most of the time, a custom object that implements only a subset--even a fairly small subset--of, e.g., the methods of a built-in dict are dictionary-like enough for purposes at hand. It would be nice, however, to be able to explicitly codify what an object needs to be able to do to be used within a given function, module, class, or framework. That's (part of) what the PyProtocols package does.

In programming languages that have static type declarations, you generally need to cast or convert data from one type to another to use it in a new context. In other languages, conversions are peformed implicitly as a context requires them, and these are called coercions. Python contains a mixture of casts and coercions, with a usual preference for the former ("explicit is better than implicit"). You can add a float to an integer, and wind up with a more general float as a result; but if you want to convert the string "3.14" into a number, you need to use the explicit constructor float("3.14").

PyProtocols provides a capability called "adaptation," which is akin to the eccentric CS concept of "partial typing." Adaptation might also be thought of as "coercion on steroids." If an interface defines a set of needed capabilities (i.e. object methods), adaptation--via the function protocols.adapt()--is a request to an object to do "whatever is necessary" to provide the needed capabilities. Obviously, if you have an explicit conversion function to turn an object of type X into one of type Y (where Y implements some interface IY), that function suffices to adapt X to protocol IY. However, adaptation in PyProtocols can do quite a bit more than this as well. For example, even if you have never explicitly programmed a conversion from type X to type Y, adapt() can often deduce a route to let X provide the capabilities mandated by IY (e.g. by finding intermediate conversions to from X to interface IZ, from IZ to IW, and from IW to IY).

Declaring Interfaces And Adaptors

There are quite a few different way within PyProtocols to create interfaces and adaptors. The PyProtocols documentation goes into these techniques in quite a bit of detail--much of which cannot be fully covered in this article. We will get into a bit of the detail below, but I think a useful approach here is to present a minimal realistic example of actual PyProtocols code.

For the example, I decided to create a Lisp-like serialization of Python objects. The representation is not literally any Lisp dialect, and I am not particularly interested in the precise pros and cons of this format. The idea here is just to create something that performs a job similar to the function repr() or the module pprint, but with a result that is both obviously different from and more easily extensible/customizable than are aforsaid serializers. One very un-Lisp-like choice was made for illustrative purposes: mappings are a more fundamental data structure than are lists (a Python tuple or list is treated as a mapping whose keys are sequential integers). Here is the code:

lispy.py PyProtocol definitions

from protocols import *
from cStringIO import StringIO


# Like unicode, & even support objects that don't explicitly support ILisp
ILisp = protocolForType(unicode, ['__repr__'], implicit=True)
# Class for interface, but no methods specifically required
class ISeq(Interface): pass
# Class for interface,  extremely simple mapping interface
class IMap(Interface):
    def items():
        "A requirement for a map is to have an .items() method"


# Define function to create an Lisp like representation of a mapping
def map2Lisp(map_, prot):
    out = StringIO()
    for k,v in map_.items():
        out.write("(%s %s) " % (adapt(k,prot), adapt(v,prot)))
    return "(MAP %s)" % out.getvalue()
# Use this func to convert an IMap-supporting obj to ILisp-supporting obj
declareAdapter(map2Lisp, provides=[ILisp], forProtocols=[IMap])
# Note that a dict implements an IMap interface with no conversion needed
declareAdapter(NO_ADAPTER_NEEDED, provides=[IMap], forTypes=[dict])

# Define and use func to adapt an InstanceType obj to the ILisp interface
from types import InstanceType
def inst2Lisp(o, p):
    return "(CLASS '(%s) %s)" % (o.__class__.__name__, adapt(o.__dict__,p))
declareAdapter(inst2Lisp, provides=[ILisp], forTypes=[InstanceType])

# Define a class to adapt an ISeq-supporting obj to an IMap-supporting obj
class SeqAsMap(object):
    advise(instancesProvide=[IMap],
           asAdapterForProtocols=[ISeq] )
    def __init__(self, seq, prot):
        self.seq = seq
        self.prot = prot
    def items(self):    # Implement the IMap required .items() method
        return enumerate(self.seq)
# Note that list, tuple implement an ISeq interface w/o conversion needed
declareAdapter(NO_ADAPTER_NEEDED, provides=[ISeq], forTypes=[list, tuple])

# Define a lambda func to adapt str, unicode to ILisp interface
declareAdapter(lambda s,p: "'(%s)" % s,
               provides=[ILisp], forTypes=[str,unicode])

# Define a class to adapt several numeric types to ILisp interface
# Return a string (ILisp-supporting) directly from instance constructor
class NumberAsLisp(object):
    advise(instancesProvide=[ILisp],
           asAdapterForTypes=[long, float, complex, bool] )
    def __new__(klass, val, proto):
        return "(%s %s)" % (val.__class__.__name__.upper(), val)

In the above code, I have declared a number of adapters in several different ways. In some cases, the code converts one interface to another interface; in other cases types themselves are directly adapted to an interface. I would like you to notice a few things about the code: (1) No adapter from a list or tuple to the ILisp interface was created; (2) No adapter is explicitly declared for the int numeric type; (3) For that matter, no adapter directly from a dict to ILisp is declared. How will the code adapt() various Python objects:

test_lispy.py object serialization

from lispy import *
from sys import stdout, stderr
toLisp = lambda o: adapt(o, ILisp)
class Foo:
    def __init__(self):
        self.a, self.b, self.c = 'a','b','c'
tests = [
  "foo bar",
  {17:2, 33:4, 'biz':'baz'},
  ["bar", ('f','o','o')],
  1.23,
  (1L, 2, 3, 4+4j),
  Foo(),
  True,
]
for test in tests:
    stdout.write(toLisp(test)+'\n')

When run, we get:

test_lispy.py serialization results

$ python2.3 test_lispy.py
'(foo bar)
(MAP (17 2) ('(biz) '(baz)) (33 4) )
(MAP (0 '(bar)) (1 (MAP (0 '(f)) (1 '(o)) (2 '(o)) )) )
(FLOAT 1.23)
(MAP (0 (LONG 1)) (1 2) (2 3) (3 (COMPLEX (4+4j))) )
(CLASS '(Foo) (MAP ('(a) '(a)) ('(c) '(c)) ('(b) '(b)) ))
(BOOL True)

Some explanation of our output would help. The first line is simple, we defined an adapter directly from a string to ILisp, the call to adapt("foo bar", ILisp) just returns the results of the lambda function. The next line is just a smidgeon more complicated. No adapter takes us directly from a dict to ILisp; but we can adapt dict to IMap without any adapter (we declared as much), and we have an adapter from IMap to ILisp. Likewise, for the later lists and tuples, we can adapt either to ISeq, ISeq to IMap, and IMap to ILisp. PyProtocols performs all the magic of figuring out what adaptation path to take, behind the scenes. A old-style instance is the same story as a string or an IMap-supporting object, we have an adaptation directly to ILisp.

But wait a moment. What about all the integers used in our dict and tuple objects? Numeric long, complex, float and bool have an explicit adapter, but int lacks any. The trick here is that an int object already has a .__repr__() method; by declaring implicit support as part of the ILisp interface, we are happy to use the existing .__repr__() method of objects as support for the ILisp interface. In particular, as a built-in, integers are represented with bare digits, rather than with a capitalized type intializer (e.g. LONG).

Adaptation Protocol

Let us look a bit more explicitly at what the function protocol.adapt() does. In our example we used the "declaration API" to implicitly setup a collection of "factories" for adaptation. This API itself has several levels to it. The "primitives" of the declaration API are the functions: declareAdaptorForType(), declareAdaptorForObject() and declareAdaptorForProtocol(). The prior example did not use these, but rather higher level APIs like: declareImplementation(), declareAdaptor(), adviceObject() and protocolForType(). In one case we saw the "magic" declaration advise() within a class body. The function advise() allows a large number of keyword arguments that configure the purpose and role of a class so advised. You may also advise() a module object.

You do not need to use the declaration API to create adapatable objects, or interfaces that know how to adapt() objects to them. Let us look at the call signature of adapt(), then explain the procedure it follows. A call to adapt() looks like:

Call signature of adapt()

adapt(component, protocol, [, default [, factory]])

What this says is that you would like to adapt the object component to the interface protocol. If default is specified, it can be returned as a wrapper object or modification for component. If factory is specified as a keyword argument, a conversion factory can be used to produce the wrapper or modification. But let us back up a little bit, and look at the complete sequence of actions attempted by adapt() (in simplified code):

Hypothetical implementation of adapt()

if isinstance(component, protocol):
    return component
elif hasattr(component,'__conform__'):
    return component.__conform__(protocol)
elif hasattr(protocol,'__adapt__'):
    return protocol.__adapt__(component)
elif default is not None:
    return default
elif factory is not None:
    return factory(component, protocol)
else:
    NotImplementedError

There are a couple qualities that calls to adapt() should maintain (but this is advise to programmers, not generally enforced by the library). Calls to adapt() should be idempotent. That is, for an object x and a protocol P, we hope that: adapt(x,P)==adapt(adapt(x,P),P). In style, this intent is similar to that behind iterator classes that return self from the .__iter__() method. Basically, you do not want re-adapting to the same thing you already adapted to to produce fluctuating results.

It is also worth noting that adaptation might be lossy. In order to bring an object into conformance with an interface, it may be inconvenient or impossible to keep all the information necessary to re-create the initial object. That is, in the general case, for object x and protocols P1 and P2: adapt(x,P1)!=adapt(adapt(adapt(x,P1),P2),P1).

Before concluding, let us look at another test script that takes some advantage of the lower level behavior of adapt():

test_lispy2.py object serialization

from lispy import *
class Bar(object):
    pass
class Baz(Bar):
    def __repr__(self):
        return "Represent a "+self.__class__.__name__+" object!"
class Bat(Baz):
    def __conform__(self, prot):
        return "Adapt "+self.__class__.__name__+" to "+repr(prot)+"!"
print adapt(Bar(), ILisp)
print adapt(Baz(), ILisp)
print adapt(Bat(), ILisp)
print adapt(adapt(Bat(), ILisp), ILisp)


$ python2.3 test_lispy2.py
<__main__.Bar object at 0x65250>
Represent a Baz object!
Adapt Bat to WeakSubset(<type 'unicode'>,('__repr__',))!
'(Adapt Bat to WeakSubset(<type 'unicode'>,('__repr__',))!)

It turns out the design of lispy.py fails the idempotence goal. A good exercise might be to improve this design. The representation as ILisp, however, is certainly lossy of the information in the original object (which is fine).

Conclusion

In feel, PyProtocols has some commonalities with other "exotic" topic this column has addressed. For one thing, the declaration API is, well, declarative (in contrast to imperative). Rather than giving the steps and switches needed to perform an action, declarative programming states that certain things hold, and lets a library or compiler figure out the details about how to carry it out. The names "declare*()" and "advice*()" are well chosen from this perspective.

As well, however, I find that PyProtocols programming has a similarity to programming with multiple dispatch, specifically with the gnosis.magic.multimethods module that I have presented in another installment. My own module performs a relatively simple deduction of relevant ancestor classes to dispatch, in contrast to PyProtocols' determination of adaptation paths. But both libraries tend to encourage a similar kind of modularity in programming--lots of little functions or classes to perform "pluggable" tasks, without necessarily falling into rigid class hierarchies. The style has an elegance to it, in my opinion.

Resources

The reference/tutorial document "Component Adaptation + Open Protocols = The PyProtocols Package" is the place to start in exploring the intricacies of PyProtocols.

http://peak.telecommunity.com/protocol_ref/ref.html

The home page for PEAK itself is the place to start for an introduction to the library as a whole.

http://peak.telecommunity.com/

And earlier Charming Python installment looked at creating declarative mini-languages within Python.

http://www-106.ibm.com/developerworks/library/l-cpdec.html

Another prior Charming Python developed and presented a library to enable multiple dispatch.

http://www-106.ibm.com/developerworks/linux/library/l-pydisp.html

Michele Simionato and I worte two articles on metaclasses, an introduction.

http://www-106.ibm.com/developerworks/linux/library/l-pymeta.html

And a look at some intracacies of metaclasses behavior.

http://www-106.ibm.com/developerworks/library/l-pymeta2/

The Twisted library is similar to PEAK, both in containing a concept of protocols and in various capabilities such as asynchronous programming and providing an application configuration framework.

Part I was "Understanding asynchronous networking."

http://www-106.ibm.com/developerworks/linux/library/l-twist1.html

Part II covered "Implementing Web servers."

http://www-106.ibm.com/developerworks/linux/library/l-twist2.html

Part III looked at "Stateful Web servers and templating."

http://www-106.ibm.com/developerworks/linux/library/l-twist3.html

And finally, Part IV, "Secure clients and servers."

http://www-106.ibm.com/developerworks/linux/library/l-twist4.html

Part of what PEAK contains is a capability for something similar to "weightless threads" that I described in two prior Charming Python installments, "Generator-based State Machines."

http://www-106.ibm.com/developerworks/linux/library/l-pygen.html

And one on "Semi-Coroutines."

http://www-106.ibm.com/developerworks/linux/library/l-pythrd.html

About The Author

Picture of Author David Mertz believes in adaptation by selective programming. David may be reached at mertz@gnosis.cx; his life pored over at http://gnosis.cx/publish/. Check out David's book Text Processing in Python (http://gnosis.cx/TPiP/).