BUG:change formatting of assert_array_almost_equal

[ElDeveloper]

assert_array_almost_equal now prints the number of digits that were
being compared in the test instead of always printing 8 (the default
for array_repr). This would lead to uninformattive error messages.

Fixes #2367

[charris]

LGTM. Needs some tests, probably in numpy/testing/tests/test_utils.py.

[ElDeveloper]

Sounds good to me, when I checked I couldn't find test cases for any of the three functions (build_err_msg, assert_array_equal and assert_array_compare) that are being modified, should I add tests for each of them? I think it only makes sense to add them for build_err_msg but I would be interested in hearing what others think.

[charris]

Some areas of Numpy are under tested. I'd say add a test build_err_msg at a minimum, and ideally the functions that call it. Looks like assert_array_almost_equal is another one. Writing tests is often more work than fixing the bug, but in this case cut and paste will probably help. The tests serve to define what the functions should do.

[ElDeveloper]

@charris that sounds good to me, I'll add those for build_err_msg and assert_array_almost_equal.

[ElDeveloper]

On a second look, I think only build_err_msg needed the tests (which I've added and I think the travis build passed). I cannot see what test cases I should add for the other functions. Do you have any pointers?

[charris]

build_err_msg is probably enough, they look pretty thorough. What I had in mind for the other functions was just making sure they called that function with the proper precision argument.

For the tests themselves, it might be easier to parse out the numbers and just compare those rather than the complete strings. Could maybe have a small function to do that.

[ElDeveloper]

Ah, now I get it, thanks! I'll try to add those.

[ElDeveloper]

@charris I've added some other tests and added a check in build_err_msg to check if it's a scalar or an ndarray. Let me know what else is needed.

[charris]

Merged, thanks @ElDeveloper.

[ElDeveloper]

Thanks for the pointers @charris.

[juliantaylor]

thanks, the now fixed behavior has always bothered me when looking at test failures logs due to rounding issues