You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Mock with spec object does not ensure method call signatures #74772
Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.
Show more details
GitHub fields:
assignee=Noneclosed_at=Nonecreated_at=<Date2017-06-07.12:04:06.898>labels= ['3.8', 'type-bug', 'library']
title='Mock with spec object does not ensure method call signatures'updated_at=<Date2019-11-07.16:48:26.232>user='https://github.com/bclau'
Mock can accept a spec object / class as argument, making sure that accessing attributes that do not exist in the spec will cause an AttributeError to be raised, but there is no guarantee that the spec's methods signatures are respected in any way. This creates the possibility to have faulty code with passing unittests and assertions.
Steps to reproduce:
>>> import mock
>>>
>>> classSomething(object):
... deffoo(self, a, b, c, d):
... pass
...
>>>
>>> m = mock.Mock(spec=Something)
>>> m.foo()
<Mock name='mock.foo()' id='140286904787240'>
>>> # TypeError should be raised, but it isn't.
...
>>> m.foo.assert_called_once_with()
Expected behaviour: It should have raised a TypeError, since the method signature is: def meth(self, a, b, c, d):
I am slightly concerned if spec should gain more responsibility than just validating attribute access which is mentioned in the docs. With the linked PR spec below would also validate the signature which is not done in Python 3.7 so this might break code for someone who only wants to validate access attribute access and not signature of the methods. It seems the PR also adds autospec argument to Mock that is currently not supported though spec and spec_set are supported.
try:
mock_foo.non_existent
except AttributeError:
print("raises AttributeError for non-existent attribute")
# 3.7
➜ cpython git:(pr_1982) python3.7 /tmp/foo.py
raises AttributeError for non-existent attribute
# With PR
➜ cpython git:(pr_1982) ./python.exe /tmp/foo.py
Traceback (most recent call last):
File "/tmp/foo.py", line 7, in <module>
mock_foo(1, 2)
File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/unittest/mock.py", line 1009, in __call__
_mock_self._mock_check_sig(*args, **kwargs)
File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/unittest/mock.py", line 103, in checksig
sig.bind(*args, **kwargs)
File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/inspect.py", line 3016, in bindreturn args[0]._bind(args[1:], kwargs)
File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/inspect.py", line 2937, in _bindraiseTypeError('too many positional arguments') fromNoneTypeError: too many positional arguments
Spec objects are currently dumb. It would be a new feature to add signature validation to them.
I think it would be a useful feature though as currently autospec sort of obsoletes spec objects whilst being more heavyweight and harder to use.
I think it would appropriate to add to 3.8 but not to 3.7.
Compatibility with existing tests is the issue. The obvious answer is a flag to turn it on/off but that adds complexity to the API.
My gut feeling is that spec objects are far more commonly used in situations where validation would be useful than a detriment.
> On 31 Mar 2019, at 18:07, Karthikeyan Singaravelan <report@bugs.python.org> wrote:
>
>
> Karthikeyan Singaravelan <tir.karthi@gmail.com> added the comment:
>
> I am slightly concerned if spec should gain more responsibility than just validating attribute access which is mentioned in the docs. With the linked PR spec below would also validate the signature which is not done in Python 3.7 so this might break code for someone who only wants to validate access attribute access and not signature of the methods. It seems the PR also adds autospec argument to Mock that is currently not supported though spec and spec_set are supported.
>
> from unittest import mock
>
> def foo(lish):
> pass
>
> mock_foo = mock.Mock(spec=foo)
> mock_foo(1, 2)
>
> try:
> mock_foo.non_existent
> except AttributeError:
> print("raises AttributeError for non-existent attribute")
>
> # 3.7
>
> ➜ cpython git:(pr_1982) python3.7 /tmp/foo.py
> raises AttributeError for non-existent attribute
>
> # With PR
>
> ➜ cpython git:(pr_1982) ./python.exe /tmp/foo.py
> Traceback (most recent call last):
> File "/tmp/foo.py", line 7, in <module>
> mock_foo(1, 2)
> File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/unittest/mock.py", line 1009, in __call__
> _mock_self._mock_check_sig(*args, **kwargs)
> File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/unittest/mock.py", line 103, in checksig
> sig.bind(*args, **kwargs)
> File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/inspect.py", line 3016, in bind
> return args[0]._bind(args[1:], kwargs)
> File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/inspect.py", line 2937, in _bind
> raise TypeError('too many positional arguments') from None
> TypeError: too many positional arguments
>
>
I'd like spec to have signature validation. I don't think the use case for "attribute validation only" (current spec behaviour) is very strong and I'd rather not add new keywords. The mock API is already too complex.
I'll take the existing PR and modify it to use spec instead of adding the new autospec.
Activity
bclau commentedon Jun 7, 2017
Mock can accept a spec object / class as argument, making sure that accessing attributes that do not exist in the spec will cause an AttributeError to be raised, but there is no guarantee that the spec's methods signatures are respected in any way. This creates the possibility to have faulty code with passing unittests and assertions.
Steps to reproduce:
Expected behaviour: It should have raised a TypeError, since the method signature is: def meth(self, a, b, c, d):
Actual behaviour: No error.
tirkarthi commentedon Mar 31, 2019
I am slightly concerned if spec should gain more responsibility than just validating attribute access which is mentioned in the docs. With the linked PR spec below would also validate the signature which is not done in Python 3.7 so this might break code for someone who only wants to validate access attribute access and not signature of the methods. It seems the PR also adds autospec argument to Mock that is currently not supported though spec and spec_set are supported.
try:
mock_foo.non_existent
except AttributeError:
print("raises AttributeError for non-existent attribute")
# 3.7
➜ cpython git:(pr_1982) python3.7 /tmp/foo.py
raises AttributeError for non-existent attribute
# With PR
voidspace commentedon Apr 1, 2019
Spec objects are currently dumb. It would be a new feature to add signature validation to them.
I think it would be a useful feature though as currently autospec sort of obsoletes spec objects whilst being more heavyweight and harder to use.
I think it would appropriate to add to 3.8 but not to 3.7.
Compatibility with existing tests is the issue. The obvious answer is a flag to turn it on/off but that adds complexity to the API.
My gut feeling is that spec objects are far more commonly used in situations where validation would be useful than a detriment.
> On 31 Mar 2019, at 18:07, Karthikeyan Singaravelan <report@bugs.python.org> wrote: > > > Karthikeyan Singaravelan <tir.karthi@gmail.com> added the comment: > > I am slightly concerned if spec should gain more responsibility than just validating attribute access which is mentioned in the docs. With the linked PR spec below would also validate the signature which is not done in Python 3.7 so this might break code for someone who only wants to validate access attribute access and not signature of the methods. It seems the PR also adds autospec argument to Mock that is currently not supported though spec and spec_set are supported. > > from unittest import mock > > def foo(lish): > pass > > mock_foo = mock.Mock(spec=foo) > mock_foo(1, 2) > > try: > mock_foo.non_existent > except AttributeError: > print("raises AttributeError for non-existent attribute") > > # 3.7 > > ➜ cpython git:(pr_1982) python3.7 /tmp/foo.py > raises AttributeError for non-existent attribute > > # With PR > > ➜ cpython git:(pr_1982) ./python.exe /tmp/foo.py > Traceback (most recent call last): > File "/tmp/foo.py", line 7, in <module> > mock_foo(1, 2) > File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/unittest/mock.py", line 1009, in __call__ > _mock_self._mock_check_sig(*args, **kwargs) > File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/unittest/mock.py", line 103, in checksig > sig.bind(*args, **kwargs) > File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/inspect.py", line 3016, in bind > return args[0]._bind(args[1:], kwargs) > File "/Users/karthikeyansingaravelan/stuff/python/cpython/Lib/inspect.py", line 2937, in _bind > raise TypeError('too many positional arguments') from None > TypeError: too many positional arguments > >voidspace commentedon Sep 10, 2019
I'd like spec to have signature validation. I don't think the use case for "attribute validation only" (current spec behaviour) is very strong and I'd rather not add new keywords. The mock API is already too complex.
I'll take the existing PR and modify it to use spec instead of adding the new autospec.
voidspace commentedon Sep 10, 2019
This will affect spec and spec_set.