name lookup failure using metaclasses with unittests

[Ulrich Eckhardt]

Hi!

I'm having problems using a metaclass to generate test functions. This
works when I try to run all tests from the module or test case, but it
fails when I'm trying to specify a single test function. My environment
is Python 2.7.3 on MS Windows 7 at the moment. It should be upgraded to
at least 2.7.4 or better to 3, but see the notes on Python 3 below.

# my_module.py
import unittest
class X(unittest.TestCase):
def __metaclass__(name, bases, dict):
# attach function
def test(self):
pass
dict['test_1'] = test
dict['test_2'] = test
# create class
return type(name, bases, dict)

The error when I'm trying to run "python -m unittest my_module.X.test_1"
is: "Value error: no such test method in <class 'my_module.X'>: test".
The astonishing part is that it claims that "test" is not found while I
asked it to run "test_1". The name it complains about is the name of the
function inside the metaclass function. In all other cases, like e.g.
giving "-v" it reports the correct function name. My question here is
whether I'm doing something wrong or whether I discovered a bug.


Now, concerning Python 3, it fails to detect any test case at all! My
guess is that the unittest library was changed to use metaclasses itself
in order to detect classes derived from unittest.TestCase. Therefore,
overriding the metaclass breaks test case discovery. My question in that
context is how do I extend metaclasses instead of overriding it? In
other words, what is the equivalent to super() for class creation?

Thank you for your help!

Uli

 

[Peter Otten]

Hi!

I'm having problems using a metaclass to generate test functions. This
works when I try to run all tests from the module or test case, but it
fails when I'm trying to specify a single test function. My environment
is Python 2.7.3 on MS Windows 7 at the moment. It should be upgraded to
at least 2.7.4 or better to 3, but see the notes on Python 3 below.

# my_module.py
import unittest
class X(unittest.TestCase):
def __metaclass__(name, bases, dict):
# attach function
def test(self):
pass
dict['test_1'] = test
dict['test_2'] = test
# create class
return type(name, bases, dict)

The error when I'm trying to run "python -m unittest my_module.X.test_1"
is: "Value error: no such test method in <class 'my_module.X'>: test".
The astonishing part is that it claims that "test" is not found while I
asked it to run "test_1". The name it complains about is the name of the
function inside the metaclass function. In all other cases, like e.g.
giving "-v" it reports the correct function name. My question here is
whether I'm doing something wrong or whether I discovered a bug.

Here's a simpler demo of the problem:

$ cat tmp.py
import unittest

class X(unittest.TestCase):
def test_1(self): pass
test_1.__name__ = "test_2"

$ python -m unittest -v tmp
test_1 (tmp.X) ... ok

----------------------------------------------------------------------
Ran 1 test in 0.001s

OK
$ python -m unittest -v tmp.X.test_1
Traceback (most recent call last):
File "/usr/lib/python2.7/runpy.py", line 162, in _run_module_as_main
"__main__", fname, loader, pkg_name)
File "/usr/lib/python2.7/runpy.py", line 72, in _run_code
exec code in run_globals
File "/usr/lib/python2.7/unittest/__main__.py", line 12, in <module>
main(module=None)
File "/usr/lib/python2.7/unittest/main.py", line 94, in __init__
self.parseArgs(argv)
File "/usr/lib/python2.7/unittest/main.py", line 149, in parseArgs
self.createTests()
File "/usr/lib/python2.7/unittest/main.py", line 158, in createTests
self.module)
File "/usr/lib/python2.7/unittest/loader.py", line 128, in
loadTestsFromNames
suites = [self.loadTestsFromName(name, module) for name in names]
File "/usr/lib/python2.7/unittest/loader.py", line 109, in
loadTestsFromName
return self.suiteClass([parent(obj.__name__)])
File "/usr/lib/python2.7/unittest/case.py", line 191, in __init__
(self.__class__, methodName))
ValueError: no such test method in <class 'tmp.X'>: test_2

It looks like this particular invocation relies on class attribute and
function __name__ being identical.

Please file a bug report.

Now, concerning Python 3, it fails to detect any test case at all! My
guess is that the unittest library was changed to use metaclasses itself
in order to detect classes derived from unittest.TestCase. Therefore,
overriding the metaclass breaks test case discovery. My question in that
context is how do I extend metaclasses instead of overriding it? In
other words, what is the equivalent to super() for class creation?

Python 3 does not recognize the __metaclass__ attribute as the metaclass.
You need to provide it like so:

def __metaclass__(name, bases, dict):
...

class X(unittest.TestCase, metaclass=__metaclass__):
pass

 

[Ulrich Eckhardt]

Am 10.04.2013 11:52, schrieb Peter Otten:

Ulrich Eckhardt wrote: [...]
It looks like this particular invocation relies on class attribute and
function __name__ being identical.

Please file a bug report.

Thanks for confirming this and reducing the test case even more.

Python 3 does not recognize the __metaclass__ attribute as the metaclass.
You need to provide it like so:

def __metaclass__(name, bases, dict):
...

class X(unittest.TestCase, metaclass=__metaclass__):
pass

:|

Doing some research[0, 1] on metaclasses in 2 and 3, I have a few more
questions...

The first thing I was wondering was why Python doesn't complain about a
class property that is marked as special (leading and trailing double
underscores) but that it knows nothing about. Worse, Python 3 should be
aware of its legacy and recognize the Python 2 metaclass syntax, even if
only to reject it loudly. I'm pretty sure there is a reason for that,

The second question that came up was if there is a way to keep a
metaclass defined inside the class or if the only way is to provide it
externally. The reason I like this in-class definition is that for my
case of autogenerated test functions, everything is in one place which
used to be in a loop that modified the class after its creation. Maybe
I'm just too brainwashed by static languages though.

To get the Python2 feeling back, I have a hack in mind that involves
creating a baseclass which in turn provides a metaclass that invokes a
specific function to post-initialize the class, similar to the way
Python 2 does it automatically, but I'm wondering if there isn't
anything better. Also PEP 3115 "Metaclasses in Python 3000"[2] seems to
consider postprocessing of a class definition as better handled by a
class decorator, which is something I haven't looked at yet.

Greetings from Hamburg!

Uli


[0] http://mikewatkins.ca/2008/11/29/python-2-and-3-metaclasses/
[1] http://www.artima.com/weblogs/viewpost.jsp?thread=236234
[2] http://www.python.org/dev/peps/pep-3115/

 

[Ulrich Eckhardt]

Am 10.04.2013 11:52, schrieb Peter Otten:

It looks like this particular invocation relies on class attribute and
function __name__ being identical.

Please file a bug report.

http://bugs.python.org/issue17696

Uli

 

[Steven D'Aprano]

The first thing I was wondering was why Python doesn't complain about a
class property that is marked as special (leading and trailing double
underscores) but that it knows nothing about.

Because that breaks backward compatibility.

You write a class in Python 2.6 or 2.7, and make it backward compatible
with 2.5:

class MyIterator(object):
def __next__(self):
...
next = __next__


Now you import it into Python 2.5, which has no idea about __next__ so it
blows up. (For no good reason, since __next__ is harmless in 2.5.)

Worse, Python 3 should be
aware of its legacy and recognize the Python 2 metaclass syntax, even if
only to reject it loudly. I'm pretty sure there is a reason for that,

That will break any future re-use of __metaclass__. It will also make
version agnostic code much harder:


class Meta(type):
...

if sys.version >= '3':
kwargs = {'metaclass': Meta}
else:
kwargs = {}

class MyClass(object, **kwargs):
__metaclass__ = Meta

The second question that came up was if there is a way to keep a
metaclass defined inside the class or if the only way is to provide it
externally. The reason I like this in-class definition is that for my
case of autogenerated test functions, everything is in one place which
used to be in a loop that modified the class after its creation. Maybe
I'm just too brainwashed by static languages though.

Not in general, since the metaclass has to exist independently of the
class.

The class is an instance of your metaclass. That means that the metaclass
must exist first, so it can be instantiated when you define the class.

However, there is a loophole: Python's metaclass machinery is actually
more general than just class-of-classes. The metaclass doesn't have to be
a class, it can be any callable with the same function signature as the
three-argument version of type. So despite what I said above, you *can*
embed the metaclass in the class, if the metaclass is a function created
with lambda (or equivalent):


# Python 2 version
class MyClass(object):
__metaclass__ = (lambda name, bases, dict:
sys.stdout.write("Spam!\n")
or type(name, bases, dict)
)


# Python 3 version
class MyClass(object, metaclass=lambda name, bases, dict:
sys.stdout.write("Spam!\n") and type(name, bases, dict)
):
pass


But really, except as a trick, why would you do that?

To get the Python2 feeling back, I have a hack in mind that involves
creating a baseclass which in turn provides a metaclass that invokes a
specific function to post-initialize the class, similar to the way
Python 2 does it automatically, but I'm wondering if there isn't
anything better.

Seems awfully convoluted and complicated. Python 3 metaclasses work
exactly the same as Python 2 metaclasses, except the syntax for declaring
them is slightly different. So if you had this:

class Meta(type):
...

class MyClass:
__metaclass__ = Meta
...


just change it to this, and it should work exactly the same:


class Meta(type):
...

class MyClass(metaclass=Meta):
...

Also PEP 3115 "Metaclasses in Python 3000"[2] seems to
consider postprocessing of a class definition as better handled by a
class decorator, which is something I haven't looked at yet.

Generally, class decorators are less brain-melting than metaclasses.

 

[Arnaud Delobelle]

The second question that came up was if there is a way to keep a metaclass
defined inside the class or if the only way is to provide it externally.

Yes, using metaclasses! I wouldn't recommend it though. Here's a
proof of concept:

class MyType(type):
def __new__(meta, name, bases, attrs):
try:
metaclass = attrs.pop('__metaclass__')
except KeyError:
return type.__new__(meta, name, bases, attrs)
else:
return metaclass(name, bases, attrs)

class MyObject(metaclass=MyType):
pass

.... def __metaclass__(name, bases, attrs):
.... print("Test metaclass")
.... return MyType(name, bases, attrs)
....
Test metaclass

 

[Ulrich Eckhardt]

Am 11.04.2013 10:19, schrieb Steven D'Aprano:

if sys.version >= '3':

Use sys.version_info >= (3,), otherwise your code breaks when upgrading
to Python 10 and greater. ;^)

The second question that came up was if there is a way to keep a
metaclass defined inside the class or if the only way is to provide it
externally. [...]

Not in general, since the metaclass has to exist independently of the
class.

Thanks for your explanations, they are appreciated.

The class is an instance of your metaclass. That means that the
metaclass must exist first, so it can be instantiated when you
define the class.

I don't like the approach to define the code to post-process a class
before defining the class. It's a bit like top-posting, it messes up the
reading order. Since I really intend to post-process the class, it seems
that metaclasses are simply not the right tool.

At the moment, this leaves me with two options:

1. post-process the class

class X:
pass
# attach constants to clas X
for i in (1, 2, 3):
setattr(X, 'f{}' % i, i)

2. generate code inline

class Y: pass
# generate constants in local (class-)namespace
for i in (1, 2, 3):
locals()['f{}' % i] = i

In both cases, variables (loop variable 'i') are leaked into the
surrounding namespace, which is kind-of ugly. The second approach also
seems a bit hackish and I can't use the class-in-definition there, which
is limiting when you want to attach e.g. constants of type X to X.

Also PEP 3115 "Metaclasses in Python 3000"[2] seems to
consider postprocessing of a class definition as better handled by a
class decorator, which is something I haven't looked at yet.

Generally, class decorators are less brain-melting than metaclasses.

Alas, they also need to be defined before the class, messing with the
mentioned order of declaration. They can be used to call a class
function though which then does the necessary postprocessing...

3. post-process the class triggered with decorator

def postprocess_class(cls):
"""invoke postprocess() on the decorated object"""
cls.postprocess()
del cls.postprocess
return cls

@postprocess_class
class Z:
@classfunction
def postprocess(cls):
# attach constants to class
for i in (1, 2, 3):
setattr(cls, 'f{}' % i, i)


I guess I'll stay with variant 1 for now, since it requires the least
amount of code and the least amount of questions from other developers here.

Thanks everybody!

Uli

 

[Terry Jan Reedy]

Am 11.04.2013 10:19, schrieb Steven D'Aprano:

if sys.version >= '3':

Use sys.version_info >= (3,), otherwise your code breaks when upgrading
to Python 10 and greater. ;^)

The second question that came up was if there is a way to keep a
metaclass defined inside the class or if the only way is to provide it
externally. [...]

Not in general, since the metaclass has to exist independently of the
class.

Thanks for your explanations, they are appreciated.

The class is an instance of your metaclass. That means that the
metaclass must exist first, so it can be instantiated when you
define the class.

I don't like the approach to define the code to post-process a class
before defining the class. It's a bit like top-posting, it messes up the
reading order. Since I really intend to post-process the class, it seems
that metaclasses are simply not the right tool.

Using a post-processing object as a metaclass or decorator necessarily
requires predefinition. Such objects are usually used more than once.

For one-off postprocessing, I probably would not bother.

At the moment, this leaves me with two options:

1. post-process the class

class X:
pass
# attach constants to clas X
for i in (1, 2, 3):
setattr(X, 'f{}' % i, i)

2. generate code inline

class Y: pass
# generate constants in local (class-)namespace
for i in (1, 2, 3):
locals()['f{}' % i] = i

Mutating class locals() currently works in CPython, but is explicitly
not guaranteed to work by the language definition.

In both cases, variables (loop variable 'i') are leaked into the
surrounding namespace, which is kind-of ugly. The second approach also
seems a bit hackish and I can't use the class-in-definition there, which
is limiting when you want to attach e.g. constants of type X to X.

Also PEP 3115 "Metaclasses in Python 3000"[2] seems to
consider postprocessing of a class definition as better handled by a
class decorator, which is something I haven't looked at yet.

Generally, class decorators are less brain-melting than metaclasses.

Alas, they also need to be defined before the class, messing with the
mentioned order of declaration. They can be used to call a class
function though which then does the necessary postprocessing...

3. post-process the class triggered with decorator

def postprocess_class(cls):
"""invoke postprocess() on the decorated object"""
cls.postprocess()
del cls.postprocess
return cls

@postprocess_class
class Z:
@classfunction
def postprocess(cls):
# attach constants to class
for i in (1, 2, 3):
setattr(cls, 'f{}' % i, i)


I guess I'll stay with variant 1 for now, since it requires the least
amount of code and the least amount of questions from other developers
here.