pytest.approx abs/rel logic seems to be flipped on np.array()

[abrammer]

Hopefully, I'm not missing something obvious here.

Absolute and relative tolerance arguments or logic seem to get flipped when the comparison is against a np.array

Think I've tracked it down to this line.

pytest/src/_pytest/python_api.py

Line 214 in f93995e

return ApproxNumpy(actual, self.abs, self.rel, self.nan_ok) == self.expected

abs and rel arguments should be in the other order according to the ApproxBase.__init__

pytest/src/_pytest/python_api.py

Line 46 in f93995e

def __init__(self, expected, rel=None, abs=None, nan_ok=False):

pytest.__version__
'3.6.3'
np.__version__
'1.13.1'

abs_check = pytest.approx(100, abs=0.1, rel=0)
rel_check = pytest.approx(100, abs=0, rel=0.1)
99 == abs_check
# False
np.array(99) == abs_check
# True

99 == rel_check
# True
np.array(99) == rel_check
# False

[pytestbot]

GitMate.io thinks possibly related issues are #3593 (pytest.approx raises TypeError for 0-d numpy array), #3312 (pytest.approx doesn't support comparison between arrays and scalars), #1994 (Supporting arrays for approx), #3482 (Extend pytest.approx to handle strings. ), and #3164 (TypeError with pytest.approx on nested dictionaries).