FIX/API: use inspect.signature for python3

[tacaswell]

This allows functions with keyword-only arguments to have their signatures introspected.

Also makes this bit of code 3.6 compatible (@Carreau I see your 3.5 and raise you a 3.6 😜 )

Found this via working on matplotlib/matplotlib#4829 where for python 3.3+ we mark the data kwarg as keyword-only.

This is a slight change in behavior as it makes all arguments, not just the optional ones. This can easily be changed back, but leaving it for now as I would propose that including all arguments is a better behavior.

[jankatins]

I wonder what would happen if on py2, this could a) unwrap (signature does unwrap until a "signature" is found, getargspec does not) and b) check for a simple get_additional_completions(...) function which we could then set in out decorator. getdoc() already does something similar...

so:

addons = []
while true:
    if hasattr(call_obj, "get_additional_completions"):
        addons.extend(call_obj.get_additional_completitions())
        break
    if hasattr(call_obj, "__wrapped__"):
        call_obj = call_obj.__wrapped__
    else:
        break
[...]
ret.extend(addons)

[minrk]

We ship a backport of inspect.signature that should work on Python 2 in IPython.utils.signatures.signature, if you want to try to just use that instead.

[jankatins]

Then mpl must ship that as well or needs a dependency on ipython...

[minrk]

@JanSchulz I'm not sure I understand. This is a PR to IPython, so why would IPython using its own utility affect MPL?

[jankatins]

This PR is part of the "unpack labeled data for plotting": make it possible to use a call like this plot("var1", "var2", data=df) instead of plot(df["var1"], df["var2"]). This is currently implemented in matplotlib/matplotlib#4829 with a decorator, which makes all signatures (= code completions and inspections in python <3.4) look like plot(*args, **kwargs). @tacaswell added a __wrapped__ attribute and a __signature__ attribute (see these two parts of the diff: https://github.com/matplotlib/matplotlib/pull/4829/files#diff-9ffb0d1db51a67ee4f37528e00715b3cR1590 https://github.com/matplotlib/matplotlib/pull/4829/files#diff-9ffb0d1db51a67ee4f37528e00715b3cR1795) in case python understands it so that newer python versions get the right signature and therefore the right code completions.

The problem is that to construct the signature object, mpl needs inspect.signature so adding that in py2.x even if IPython understands that, we would need inspect.signature included in mpl as well. Using a workaround like get_addon_completions() would not.

[tacaswell]

@JanSchulz Just moving that file over is probably a non-starter for now as mpl (still) supports py2.6 so we don't have OrderedDict.

[tacaswell]

Probably the best route on the mpl side is to try to import IPython and use it if available. I think this is ok because the only reason we want to tack the signatures on to legacy python is for nice introspection in IPython.

[jankatins]

But then your PR here should also try to use the internal ipython copy oft inspect (if python doesn't have it...) in all versions...

edit: Ah, you already did it :-)

[tacaswell]

no, mpl will only fall back to the IPython version in versions of python which don't already have signature.

[tacaswell]

@JanSchulz See jankatins/matplotlib#4

[tacaswell]

@minrk This look good?

Can this also be back-ported to what ever branch you are going to release off of next? This is needed to make a feature in matplotlib 1.5 that @fperez pushed for to work well.

[minrk]

@tacaswell the next release will be 4.1, so this will get in. We haven't started making breaking changes on IPython to push master to 5.x and start backporting, yet.