There is a little awkward thing. For fun large fun integer values (or if large values are given to the allowed difference), the behaviour would change. That should be fine, but the change needs to be documented in the function and release notes. Also, you need to channel objects through these, too; maybe it would be better as a negative list rather than a postive one.

I think the correct fix is to make allclose raise an error if non-numeric data types are passed in.

My use-case is that of record arrays.

With the patch, I get

>>> from numpy import *
>>> x = array([('a', 5)], dtype=[('x', 'S1'), ('y', 'l')])
>>> allclose(x, x)
True

without the patch, this produces an error.

I agree that the behavior for ints should not change. Maybe just change the test to .. not in 'fic'?

I am +/-0 on integer types, and OK with strings being passed through. But your example makes me more cautious. What if the record array has an inexact field? If it is possible to have an inexact number not being matched within tolerance with the function, I prefer to not change it.

@seberg - here is what I see with the patch:

>>> x = array([('a', 5)], dtype=[('x', 'S1'), ('y', 'f')])
>>> y = array([('a', 5+1e-10)], dtype=[('x', 'S1'), ('y', 'f')])
>>> allclose(x,y)
True

This is what I want: numeric fields compared with tolerance and non-numeric ones compared exactly.

@abalkin, that is because 5+1e-10 is 5 for floats, use doubles :) ('f' is float).

My bad. It looks like record arrays need to be special cased.

Let's take this in steps:

1. Support non-numeric and non-bool dtypes by doing exact match.
2. Support record arrays.
3. Support bool without relying on boolean subtraction.

On #3, I think we need to preserve the following behavior:

>>> allclose(True, False, atol=2)
True

Allclose now converts its arguments to inexact types before comparison, so I think this PR goes by the way. I think @pv suggestion to not handle non-numeric types is a good one. @abalkin For record arrays you may want to put a different function together to do what you want.