It's a feature, not a bug ;) The default print formatting will change in 1.14. We are aware that this will break doctests. From the release notes

Changes to array printing, and the new "legacy" printing mode.
--------------------------------------------------------------
The new ``sign='-'`` option (see new features below) causes the ``repr`` of
float arrays to often omit a whitespace previously printed in the sign
position, and 0d arrays also now print with small whitespace and precision
differences (see changes below). These changes are likely to break downstream
user's doctests.

These new behaviors can be disabled to mostly reproduce numpy 1.13 behavior by
enabling the new "legacy" printing mode. This is enabled by calling
``np.set_printoptions(legacy=True)``, or using the new ``legacy`` argument
to ``np.array2string``.

Note that there are various ways to relax whitespace sensitivity of doctests which are probably worth looking into if you want to continue to rely on doctests. @ahaldane Further suggestions?

Note that there are more changes than you show, for instance

In [1]: print(repr(np.array([1., 2., 3.])))
array([1., 2., 3.])

Note missing spaces.

Oops, I see that you did use the legacy mode, so we can maybe fix that.

Oops, I see that you did use the legacy mode, so we can maybe fix that.

That would be great!

For now our strategy is to use the legacy mode to have compatibility of our doctest with numpy 1.14dev and 1.13. When numpy 1.14 is released we will revisit our strategy. One possible solution is to update all of our doctests to work with 1.14 and just skip them for numpy <= 1.13.

I'm confused. So did numpy decide to definitely break, regardless of the sign/legacy setting?

(this was not the issue I meant to comment on, and this thread actually answered my question, sorry for the noise)

Something that was printed as 1.0 is now printed as 0.99999999(...?) is that also a feature?

What code exactly printed 0.99999999? There is a new float printing algorithm in 1.14 which, barring a bug, should print just like a python3 float.

I'll take a look at the space in the legacy option, I'm pretty sure that is a fixable bug.

What code exactly printed 0.99999999? There is a new float printing algorithm in 1.14 which, barring a bug, should print just like a python3 float.

There is a case for np.set_printoptions(legacy=True) to implement the pre-dragon4 printing. I am not sure it is worth the cost, I'll open a separate issue to discuss this. FWIW in scikit-learn we have managed to work around these issues by using doctest ellipsis.

Yes, we can discuss that. But if you see a number that is printed as 0.99999999 when it should be printed as 1.0 let us know.

The goal of the new algorithm is to round "correctly" for all float sizes, so if there is an instance where it isn't rounding correctly, that is important.

Just a quick update, it is now legacy='1.13' to get the old printing.