num_regression outputs garbage to its file if arrays are of different size

If you write arrays of different sizes on the data dict, num_regression fills the missing positions with garbage (or at least it seems garbage):

def testNumRegression(num_regression):
    data = {}
    for index in range(4):
        data[f'data_{index}'] = np.array([index])

    data['other'] = np.array([10, 20, 30, 40, 50, 60])

    num_regression.check(data)

Output:

,data_0,data_1,data_2,data_3,other
0,0,1,2,3,10
1,-2147483648,-2147483648,-2147483648,-2147483648,20
2,-2147483648,-2147483648,-2147483648,-2147483648,30
3,-2147483648,-2147483648,-2147483648,-2147483648,40
4,-2147483648,-2147483648,-2147483648,-2147483648,50
5,-2147483648,-2147483648,-2147483648,-2147483648,60

Output converted to a markdown table so it’s easier to visualize:

	data_0	data_1	data_2	data_3	other
0	0	1	2	3	10
1	-2147483648	-2147483648	-2147483648	-2147483648	20
2	-2147483648	-2147483648	-2147483648	-2147483648	30
3	-2147483648	-2147483648	-2147483648	-2147483648	40
4	-2147483648	-2147483648	-2147483648	-2147483648	50
5	-2147483648	-2147483648	-2147483648	-2147483648	60

I can think of some ways of solving this:

1. Raise an error suggesting that the user write the output on different files because the arrays have different sizes.
2. Separate it in two tables, but I don’t know how difficult would be for num_regression to process multiple tables on the same file.
3. Fill the missing items with the space character.

[tadeu]

With respect to the flag fill_different_shape_with_nan=True, should we raise an error when the array type is not float, and fill it with nan when the array type is float?

[tarcisiofischer]

Sounds good to me 😃