API: consider better decorator option

[chris-b1]

It's not an issue in py3.5, but in older versions this is annoying - because a decorator is used in the definition, the function signature on to_datetime is lost. This is the big one I've run into, but there may be some other places in the API it comes up.

I saw this blog today - could use one of the strategies there. It may make sense to vendor rather than adding another dependency?
https://hynek.me/articles/decorators/

[jreback]

it should propagate

I think it just needs wraps

[jreback]

actually @chris-b1 where is this from? ? in ipython and help work just fine. can you elaborate.

[jreback]

actually read that article. Ok that is nice.

[chris-b1]

I'm assuming you're looking on python3? In python 2 the standard wraps doesn't propagate signatures.

In [1]: import functools

In [2]: def dec_stdlib(f):
   ...:     @functools.wraps(f)
   ...:     def inner(*args, **kwargs):
   ...:         f(*args, **kwargs)
   ...:     return inner

In [6]: import boltons.funcutils

In [7]: def dec_boltons(f):
   ...:     @boltons.funcutils.wraps(f)
   ...:     def inner(*args, **kwargs):
   ...:         f(*args, **kwargs)
   ...:     return inner

In [4]: @dec_stdlib
   ...: def f(a, b):
   ...:     return a + b

In [5]: f?
Signature: f(*args, **kwargs)
Docstring: <no docstring>
Type:      function


In [8]: @dec_boltons
   ...: def f(a, b):
   ...:     return a + b

In [9]: f?
Signature: f(a, b)
Type:      function

[chris-b1]

I attempted this using https://github.com/mahmoud/boltons/blob/master/boltons/funcutils.py - which has a wraps replacement, but the interaction of the magic broke a bunch of stuff. I'm sure it would be possible to work through it, but I'm letting this one sit - I guess it's good self-motivation to update things to python 3.

[topper-123]

I've looked into this a bit using python 3.7 and the getfullargspecs problem still stands in 3.7. I'll remove the 2/3 compat label, but let the issue stand open.

For the record, giving the drawbacks to wrapt and decorators.py, I like vendoring boltons.functools.wraps the best (assuming it can be pulled out painlessly - haven't checked that).
.

[jbrockmendel]

@chris-b1 since we're now py35+ is this closeable?

[chris-b1]

I would think so! Although I am not following @topper-123 comment above on getfullargspecs so maybe some part of this that doesn't work?

[topper-123]

My point was that inspect.getfullargspec still does not work 100 %. Quoting the the linked article:

It’s kind of cheating: the signatures are still broken. inspect.signature() is just smart enough to lie to you.

OTOH, if mypy accepts the "lie", all the better. Does anyone know? (PyCharm does not).

To demonstrate the difference between a decorated and non-decorated function:

>>> from inspect import getfullargspec
>>> from functools import wraps
>>> def my_func(a:int, b: int = 1) -> int:
    ...:     return a + b
>>> def deco(func):
...     @functools.wraps(func)
...     def inner_func(*args, **kwargs):
...         return func(*args, **kwargs)
...     return inner_func

>>> getfullargspec(my_func)
FullArgSpec(args=['a', 'b'], varargs=None, varkw=None, defaults=(1,), kwonlyargs=[], kwonlydefaults=None, annotations={'return': <class 'pandas.core.series.Series'>, 'a': <class 'pandas.core.series.Series'>, 'b': <class 'int'>})
>>> getfullargspec(deco(my_func))
FullArgSpec(args=[], varargs='args', varkw='kwargs', defaults=None, kwonlyargs=[], kwonlydefaults=None, annotations={'return': <class 'pandas.core.series.Series'>})

Notice the output is quite different.

inspect.signature gives the same output for both decorared and non-decorated, but that's not sufficient for at least PyCharm to treat a decorated function as the same as the non-decorated.

[jbrockmendel]

@topper-123 now that we're on the verge of dropping py35, does that affect whether this is closeable and/or actionable?

[jbrockmendel]

@WillAyd there was another decorator-related issue that just got closed. can this one go too?

[WillAyd]

I think this works on master?

In [3]: inspect.getfullargspec(pd.to_datetime)                                                                                                                                                        
Out[3]: FullArgSpec(args=['arg', 'errors', 'dayfirst', 'yearfirst', 'utc', 'format', 'exact', 'unit', 'infer_datetime_format', 'origin', 'cache'], varargs=None, varkw=None, defaults=('raise', False, False, None, None, True, None, False, 'unix', True), kwonlyargs=[], kwonlydefaults=None, annotations={'return': typing.Union[pandas.core.indexes.datetimes.DatetimeIndex, ForwardRef('Series'), ~DatetimeScalar, ForwardRef('NaTType')], 'arg': typing.Union[~DatetimeScalar, typing.List, typing.Tuple, ~ArrayLike, ForwardRef('Series')], 'errors': <class 'str'>, 'dayfirst': <class 'bool'>, 'yearfirst': <class 'bool'>, 'utc': typing.Union[bool, NoneType], 'format': typing.Union[str, NoneType], 'exact': <class 'bool'>, 'unit': typing.Union[str, NoneType], 'infer_datetime_format': <class 'bool'>, 'cache': <class 'bool'>})

So sure let's close - can always reopen