assert_allclose reports NaN as max relative difference if arrays have zeros at the same positions

[CakeWithSteak]

If two different arrays share a zero at the same position the max relative difference will be reported as NaN by assert_allclose. This is because in assert_array_compare the error is simply calculated as max_rel_error = (error / abs(y)).max() which can result in 0/0 = NaN.
This hides the actual biggest relative difference and makes it more difficult to solve the problem in your code that is triggering the assertion.

Reproducing code example:

import numpy as np
x = np.array([0, 1])
y = np.array([0, 2])
np.testing.assert_allclose(x, y, rtol=0.001)

Output:

AssertionError: 
Not equal to tolerance rtol=0.001, atol=0

Mismatch: 50%
Max absolute difference: 1
Max relative difference: nan
 x: array([0, 1])
 y: array([0, 2])

Expected output:

AssertionError: 
Not equal to tolerance rtol=0.001, atol=0

Mismatch: 50%
Max absolute difference: 1
Max relative difference: 0.5
 x: array([0, 1])
 y: array([0, 2])

Numpy/Python version information:

1.16.4