Specifications: Contracts for Busy Developers

The QP package has a wonderful module, qp.lib.spec which deserves to get more attention whether QP does or not.

spec provides an easy way to declare or specify what various object attributes should contain, without littering your code with miles of asserts and other test. Examples will make things clear - lets take our too-simple Entry object and beef it up. First the original database-aware object:

from durus.persistent import PersistentObject

class Entry(PersistentObject):
    title = None
    text = None
    author = None
    created = None

Out of control: Clearly the object as its described presents a problem - its just an empty 'bag' with no constraints on what a user/developer might attempt to do with it. For example, whether intended or not, our dumb object allows for all sorts of questionable attribute assignments, as shown in this interactive session:

->> e = Entry()
->> e.title = 'my title'  # so far so good
->> e.title = None        # probably reasonable
->> e.title = 123.456     # not at all what we want
->> e.created = 3.1415
->> e.some_new_attribute = "foo"

The dynamic nature of Python is both blessing and curse at times; the above object requires lots of additional code to ensure that data it manages is what the developer intended, leading to significant code expansion via tests and assertions, reducing readability along the way.

In control: There is another way. Lets use spec and apply specifications and some helper methods:

from durus.persistent import PersistentObject
from qp.lib.spec import datetime_with_tz, spec, string
from qp.lib.spec import add_getters_and_setters

class Entry(PersistentObject):

    title_is = spec(
        (string, None),
        'A short description of the entry')
    text_is = spec(
        (string, None),
        'The full text of the entry')
    author_is = spec(
        User,
        'The individual responsible for the content of the entry')
    created_is = datetime_with_tz

add_getters_and_setters(Entry)

A spec can be a simple type assignment (see created_is above) or make use of the spec function which allows for a certain amount of self-documentation which can be very helpful. Arguments supplied in tuple imply the specification either, or you can spell it out: either(string, None).

Now lets run through the same interactive session as before:

->> e = Entry()
->> e.set_title('my title')
->> e.set_title(None)
->> e.set_title(123.456)
Traceback (most recent call last):
  File "<input>", line 2, in <module>
  File "/usr/local/lib/python2.5/site-packages/qp/lib/spec.py", line 725, in f
    require(value, getattr(klass, name + '_is'))
  File "/usr/local/lib/python2.5/site-packages/qp/lib/spec.py", line 171, in require
    raise TypeError(error)
TypeError:
  Expected: (string, None)
  A short description of the entry
  Got: 123.456

Aha, for the cost of a pair of get and set methods (no moaning please, we didn't even have to write the getter and setter ourselves and they don't clutter the code), we've effectively constrained what type of data can be assigned to the title attribute of Entry, while also preserving the easy to read nature of the original, simple, code.

To demonstrate the utility of spec, I've pulled a number of examples from Dulcinea and my own code. As you can see, there is great flexibility provided:

date_is = datetime

approvals_is = spec(
    sequence(DulcineaUser, set),
    "The users who agree the issue is resolved.")

issues_is = spec(
    mapping({string:Issue}, PersistentDict),
    "Mapping of issue IDs to issues.")

id_is = spec(
    pattern('^[-A-Za-z0-9_@.]*$'),
    "unique among users here")

# some specs are referred to over and over again

datetime_without_tz = both(datetime, with_attribute(tzinfo=None))
datetime_with_tz = both(datetime, with_attribute(tzinfo=no(None)))
email_pattern = pattern("^.+@.+\..{2,4}$")
existing_user = both(User, with_attribute(id=no(None)))
hex_pattern = pattern('[a-fA-F0-9]*$')

# reuse them

email_is = email_pattern
id_is = spec(
    hex_pattern,
    'a lower case alphanumeric pattern')

# use the specifications in tests (more powerful and cleaner than
# testing for type and instance alone)

require(thing, either(list, tuple))
match(a_user, existing_user) # returns boolean

My experience is that you can create fairly complex specifications that remain very readable. Have a complex object that has to be "just so" before it is committed to a database in a transaction? Specify everything, and check it for sanity with a one line assertion: assert get_spec_problems(theobject_instance) == [] and you are done.

Testing, Testing, One Two Three

As you might imagine, it becomes easier to write unit tests when our objects are so highly specified. Sancho, a unit testing framework also from the same development shop from which QP originates, is designed for projects and teams who prefer to leave code in a working state, all or most of the time.

Tests live in ./test, one level down from our objects being tested, and there is no __init__.py. A utility, urun.py, will execute one test supplied on the command line, or all tests in the test subdirectories in the current directory and below. Lets write one for Entry:

# /www/lib/parlez/test/utest_journal.py
from parlez.journal import Entry
from sancho.utest import UTest, raises

entry_text = '''This is a blog entry.\n\n*We hope you like it*.'''

class EntryTest(UTest):

    def init_test(self):
        Entry()

    def entry_test(self):
        joe = User('joe')
        e = Entry()
        e.set_author(joe)
        # a string causes a TypeError, authors must be User instances
        raises(TypeError, e.set_author, 'Joe')
        assert e.get_author() == joe
        assert e.get_created() == e.get_stamp()
        e.set_text(entry_text)
        assert e.get_text() == entry_text
        e.set_stamp()
        assert e.get_created() != e.get_stamp()

class JournalTest(UTest):
    # we'll write this shortly, before Journal!
    pass

if __name__ == '__main__':
    EntryTest()
    JournalTest()

Run urun.py from the command line or from your editor and the result:

# /www/lib/parlez/test% urun.py
./utest_journal.py: EntryTest:

No tracebacks indicates successful test(s).

In part five of this series we'll start to write the HTML (remember, this article series is apparently about web development with QP) and other user interfaces for our Entry object.