4.0.x: 'ValueError: Sentinels must not start with _' and others
2uasimojo opened this issue · comments
With 4.0.0 and 4.0.1, tests start failing in several OpenStack projects. Example 1:
b'Traceback (most recent call last):'
b' File "/home/efried/openstack/nova/nova/tests/unit/virt/libvirt/test_host.py", line 589, in test_cpu_features_bug_1217630'
b" with mock.patch('nova.virt.libvirt.host.libvirt') as mock_libvirt:"
b' File "/home/efried/openstack/nova/.tox/py37/lib/python3.7/site-packages/oslotest/mock_fixture.py", line 171, in __enter__'
b' _lazy_autospec_method(mocked_method, original_attr, eat_self)'
b' File "/home/efried/openstack/nova/.tox/py37/lib/python3.7/site-packages/oslotest/mock_fixture.py", line 27, in _lazy_autospec_method'
b' _lazy_autospec = mock.create_autospec(original_method)'
b' File "/home/efried/openstack/nova/.tox/py37/lib/python3.7/site-packages/mock/mock.py", line 2679, in create_autospec'
b' name=_name, **_kwargs)'
b' File "/home/efried/openstack/nova/.tox/py37/lib/python3.7/site-packages/mock/mock.py", line 2076, in __init__'
b' _safe_super(MagicMixin, self).__init__(*args, **kw)'
b' File "/home/efried/openstack/nova/.tox/py37/lib/python3.7/site-packages/mock/mock.py", line 439, in __init__'
b' self._mock_add_spec(spec, spec_set, _spec_as_instance, _eat_self)'
b' File "/home/efried/openstack/nova/.tox/py37/lib/python3.7/site-packages/mock/mock.py", line 494, in _mock_add_spec'
b' if iscoroutinefunction(getattr(spec, attr, None)):'
b' File "/home/efried/openstack/nova/.tox/py37/lib/python3.7/site-packages/mock/backports.py", line 34, in iscoroutinefunction'
b" getattr(obj, '_is_coroutine', None) is _is_coroutine"
b' File "/home/efried/openstack/nova/.tox/py37/lib/python3.7/site-packages/oslo_utils/fixture.py", line 82, in __getattr__'
b" raise ValueError('Sentinels must not start with _')"
b'ValueError: Sentinels must not start with _'
(permalink to that source line: https://opendev.org/openstack/nova/src/commit/e69dbfa0d34d6b3f51282ac0ab51cdab3c2115e4/nova/tests/unit/virt/libvirt/test_host.py#L589)
Switching from import mock to from unittest import mock fixes this one.
We will add comments with other examples.
Example 2:
b"2020-02-17 15:51:28,040 INFO [nova.console.websocketproxy] handler exception: Expected int or long, got <class 'mock.mock._AutospecMagicMock'>"
...
b'Traceback (most recent call last):'
b' File "/home/efried/openstack/nova/nova/tests/unit/console/test_websocketproxy.py", line 627, in test_tcp_rst_no_compute_rpcapi'
b' self.assertIsNone(self.wh._compute_rpcapi)'
b' File "/home/efried/openstack/nova/.tox/py37/lib/python3.7/site-packages/testtools/testcase.py", line 426, in assertIsNone'
b' self.assertThat(observed, matcher, message)'
b' File "/home/efried/openstack/nova/.tox/py37/lib/python3.7/site-packages/testtools/testcase.py", line 498, in assertThat'
b' raise mismatch_error'
b'testtools.matchers._impl.MismatchError: <nova.compute.rpcapi.ComputeAPI object at 0x7f865ddfa710> is not None'
We are seeing something similar:
Captured traceback:
Traceback (most recent call last):
File "/home/zuul/src/opendev.org/openstack/octavia/octavia/tests/unit/certificates/manager/test_barbican_legacy.py", line 81, in setUp
self.secret1 = mock.Mock(spec=secrets.Secret)
File "/home/zuul/src/opendev.org/openstack/octavia/.tox/py36/lib/python3.6/site-packages/mock/mock.py", line 1084, in __init__
_spec_state, _new_name, _new_parent, **kwargs
File "/home/zuul/src/opendev.org/openstack/octavia/.tox/py36/lib/python3.6/site-packages/mock/mock.py", line 439, in __init__
self._mock_add_spec(spec, spec_set, _spec_as_instance, _eat_self)
File "/home/zuul/src/opendev.org/openstack/octavia/.tox/py36/lib/python3.6/site-packages/mock/mock.py", line 494, in _mock_add_spec
if iscoroutinefunction(getattr(spec, attr, None)):
File "/home/zuul/src/opendev.org/openstack/octavia/.tox/py36/lib/python3.6/site-packages/mock/mock.py", line 2899, in __get__
return self()
File "/home/zuul/src/opendev.org/openstack/octavia/.tox/py36/lib/python3.6/site-packages/mock/mock.py", line 1100, in __call__
return _mock_self._mock_call(*args, **kwargs)
File "/home/zuul/src/opendev.org/openstack/octavia/.tox/py36/lib/python3.6/site-packages/mock/mock.py", line 1104, in _mock_call
return _mock_self._execute_mock_call(*args, **kwargs)
File "/home/zuul/src/opendev.org/openstack/octavia/.tox/py36/lib/python3.6/site-packages/mock/mock.py", line 1161, in _execute_mock_call
raise effect
ValueError
@2uasimojo Please add a note on the version of Python. mock 4.0 added support to detect coroutines to return AsyncMock. It's trying to detect it by getting _is_coroutine. oslotest tries to do autospec and has a restriction that getattr cannot be used with _is_coroutine. From the traceback if you are on Python 3.7 this is something that might affect your upgrade to 3.8 too. So in one way I see at as an upstream issue that is detected by the backport.
@johnsom Please add a simplified reproducer for the issue on how to reproduce the failure. From the traceback yours seem to be a different issue. Thanks.
The issue reported by @johnsom can be reproduced as below. Please correct me if I understood the issue incorrectly.
from unittest.mock import Mock, PropertyMock
class Foo:
prop = PropertyMock(side_effect=ValueError)
mock = Mock(spec=Foo)
In upstream cpython mock with 3.7 there is no error. In Python 3.8 and master it triggers ValueError. The cause in both issues reported is that we rely on getattr to detect coroutine that might have side effects that were not present before. Both will be upstream issues when the projects upgrade to 3.8. I am not sure of a way to get the attribute without triggering the side effect of getattr that could possibly solve the issue.
@2uasimojo / @johnsom - thanks for the reports, but these are separate issues and would have been better filed as such.
@tirkarthi - thanks for the digging! Sounds like these are all upstream issues and so will be need to be reported and fixed there so they can be backported.
@2uasimojo / @johnsom - please can you report these issues on https://bugs.python.org/, and file each one separately. Feel free to ping the bpo urls here when you have them, and and PRs, so that I can get a new backport release out as soon as the fixes land on cpython master.
I think the original issue reported here ('ValueError: Sentinels must not start with _') caused by the violation of the getattr protocol in oslo utils [1]. I filed a bug in oslo utils [2] and bushed a fix [3].
[1] https://github.com/openstack/oslo.utils/blob/b9938230f992935e8332b6e288937be890724cd2/oslo_utils/fixture.py#L82
[2] https://bugs.launchpad.net/oslo.utils/+bug/1885281
[3] https://review.opendev.org/#/c/738207/