Testdata attrs that are referenced in a classmethod will have different values at test runtime

[wolfadactyl]

The issue description here is admittedly terrible, so here is an example of a test that will fail to make it clearer:

class TestFoo(TestCase):
    @classmethod
    @wrap_testdata
    def setUpTestData(cls):
        super().setUpTestData()
        cls.foo = 'bar'

    @classmethod
    def validate(cls):
        assert cls.foo == 'baz'

    def test_baz(self):
        self.foo = 'baz'
        self.validate()

There are times when a classmethod makes sense - e.g. a classmethod can be called in setUpTestData whereas an instance method cannot. While this implementation is using a basic string attr, it's more critical in the case of an object attr where we may want to call that classmethod again at test runtime, however with wrap_testdata the value of the attr is the original state that was set in setUpTestData and doesn't include the changes made in the test itself.

We could refresh the value in the classmethod, however that would break wrap_testdata as now the class will have the modified version from the test going forward.

[charettes]

however with wrap_testdata the value of the attr is the original state that was set in setUpTestData and doesn't include the changes made in the test itself.

That's just how attribute assignment works in Python, if you do assert self.foo == 'baz' in test_baz you'll notice that the assignment worked properly.

class cls:
    foo = 'bar'

assert cls.foo == 'bar'
self = cls()
assert self.foo == 'bar'  # Since self.__dict__['foo']  is missing it falls back to cls.__dict__['foo']
self.foo = 'baz'  # Assigns self.__dict__['foo'] = 'baz' and leaves cls.__dict__['foo']  untouched
assert cls.foo == 'bar'
assert self.foo == 'baz'

In the example you provided, even with the decorator removed, the attribute assignment flow is the following:

1. setUpTestData; TestFoo.__dict__['foo'] = 'bar'
2. test_baz; <instance TestFoo>.__dict__['foo'] = 'baz'
3. vaidate; assert TestFoo.__dict__['foo'] == 'baz' which fails

In other words, doing self.foo = 'baz' doesn't alter TestFoo.foo in any way (that's just how class instance attribute assignment works in Python) and the testdata descriptor won't interfere with that at all.