Note that this mangling only occurs in python 2 - any test using signature would pass in python 3, since functools.wraps copies __signature__.

If we actually care about py2 signatures, then every use of functools.wraps in the codebase is "wrong".

One important question is what version of python the docs are built with - clearly it ought to be py3 if some of our signatures are missing

@eric-wieser @rgommers Ralf, what version of Python do you build the docs with?

Ralf, what version of Python do you build the docs with?

2.7. But it should work with all supported Python versions

Re: whether other decorators are broken in numpy

I only see a single other use of wraps in numpy, which was recently added in #8996, which in that case gives a pretty sensible "new" signature which I don't think needs to change.

Also, the signatures break in python 3 for py <= 3.3 too. (Though I doubt many people use those versions any more).

As far as I can tell from setup.py, it seems that numpy is not targeting 3.3 any more, so that doesn't actually matter.

But yeah, if this is the only place we run into this problem, I think you're right to fix this on all python versions by removing the decorator.

typo fixed

@eric-wieser Good to go?

Tempted to say we should use a context manager instead here, so that it can still be factored out:

@contextlib.contextlibmanager
def repr_guard(func, obj, repr_running={}):
    key = func, id(obj), get_ident()
    repeated = key in repr_running
    repr_running.add(key)
    try:
        yield repeated
    finally:
        repr_running.discard(key)

def array2string(a, ...):
    with repr_guard(array2string, a) as recursed:
        if rescursed:
            return '...'
        do_the_normal_stuff

Seems like potentially a premature generalization, but maybe that's just my bias - I like to avoid abstractions if there's no code duplication.

Is there anywhere else in numpy we might want to use this?

Nope, this is probably not going to be used anywhere else. I tend to be biased the other way, preferring abstractions even if they only get used once. It is a bit of a clumsy generalization though, since decorators weren't an option.

Since this is just a matter of opinion, I'll just go ahead and merge what you have.

Wait, actually there's a much easier fix here - keep the decorator, but move it to the internal _array2string, which doesn't need to preserve its signature

Hmm that should work.. I'll give it a try.

Done, I'm happy doing as you said. So this PR just comes down to moving _array2string.

Thanks!