BUG: Timedelta * Series -> TypeError

[shoyer]

I encountered this today when rerunning a script that worked prior to the 0.15 arrival of Timedelta:

In [2]: pd.to_timedelta(10, unit='D') * pd.Series(range(10))
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-2-a5e3bc3be480> in <module>()
----> 1 pd.to_timedelta(10, unit='D') * pd.Series(range(10))

/Users/shoyer/dev/pandas/pandas/tslib.so in pandas.tslib.Timedelta.__mul__ (pandas/tslib.c:31355)()

TypeError: cannot multiply a Timedelta with <class 'pandas.core.series.Series'>

@jreback I think there might be another issue for this already?

[aimboden]

Note that reversing the order does work:

pd.Series(range(10))*pd.to_timedelta(10, unit='D')
Out[158]: 
0    0 days
1   10 days
2   20 days
3   30 days
4   40 days
5   50 days
6   60 days
7   70 days
8   80 days
9   90 days
dtype: timedelta64[ns]

[jreback]

I recall that their was no easy way to fix this
(though will look again)
because TimeDelta is a scalar and Python thinks for some reason doesn't call rmul

[shoyer]

@jreback I can also take a look, I've written a lot of classes that do arithmetic at this point. I'm pretty sure this is fixable.

[jreback]

oh, this is a buggie. It is intercepted in cython. I don't allow any multiplication of a scalar timedelta with other objects (nothing makes sense). Except for the other way around (e.g. Series/Timedelta) if its an integer Series/Index it makes sense to allow multiple and does (so this is the bug).

I think division is right, but maybe check that too. E.g. you can only divide a Timedelta (object/series/index) by a Timedelta object (though I think allow integer division too).

[immerrr]

I don't allow any multiplication of a scalar timedelta with other objects (nothing makes sense)

This doesn't feel right.

I mean, set of possible multiplicand types is inherently open, you cannot realistically say that nothing in the infinite set of possibilities makes sense. Which is why python requires NotImplemented to be returned from binary operators so that the second operand has a shot at performing a reflected operation.

[jreback]

sorry, misspoke, numpy array/index/series of integers works. (other return a TypeError). could certainly be NotImplemetedError though).

@immerrr not much actually makes sense. Maybe you have something that does that was missed?

[immerrr]

First, it's not raise NotImplementedError, it's return NotImplemented: https://docs.python.org/2/reference/datamodel.html#object.__radd__. NotImplemented itself is documented poorly, but hopefully this will be fixed soon there's an issue that has landed recently that changed the doc to unambiguously state the following :

If one of those methods does not support the operation with the supplied arguments, it should return NotImplemented.

Restricting the set of operand types in python's world of duck typing doesn't make sense to me. For the very least think of all the possible types of ndarray-like containers that don't pass the isinstance(ndarray) check, e.g. scipy sparse matrices or xray.DataArray.

[jreback]

@immerrr
of course restricting makes sense. If only to preserve user sanity.
You don't want to constantly gave to check for NotImplemented types. Its just counter-productive.

xray can feel free to handle any cases it wants. but 99% of users want to be informed that
pd.Timedelta(....)*pd.Timedelta(..) just does not make sense

[immerrr]

You don't want to constantly gave to check for NotImplemented types. Its just counter-productive.

Could you rephrase that? I suppose you mean that you'll have to check for NotImplemented in user-level code, it's not true: python interpreter will raise a TypeError on its own if both a.__mul__(b) and b.__rmul__(a) return NotImplemented.

xray can feel free to handle any cases it wants. but 99% of users want to be informed that
pd.Timedelta(....)*pd.Timedelta(..) just does not make sense

It is exactly my point that it won't be able to do that. timedelta.__mul__ raising a TypeError preemptively won't give dataarray.__rmul__ a chance to handle the operation.

[jreback]

ok, i'll rephrase. What I mean is for 99% of the users a TypeError is the expectation for an operation that is not 'right' and NOT NotImplementedError returned / raise, which means that it just has not been done. TypeError means this does not make sense (in pandas).

If xray wants to implement then they can do it (though I suspect since its much easier to simply push this upstream, e.g. to get @shoyer to convince us that its a valid operation and support it mainstream in pandas). E.g. Interval

this is much easier IMHO that e.g. pushing upstream to numpy (TL;DR) the other discussions that we have had.

[immerrr]

ok, i'll rephrase. What I mean is for 99% of the users a TypeError is the expectation for an operation that is not 'right' and NOT NotImplementedError returned / raise, which means that it just has not been done. TypeError means this does not make sense (in pandas).

1. There may be (probably, are already) many projects that implement their own ndarray-like containers.
2. There's a non-formalized "convincing" operation which may or may not succeed or in which a certain project maintainer may or may not be interested.
3. It's not NotImplementedError, it's a singleton object called NotImplemented, and it doesn't mean "it just has not been done", it means, according to the doc, that it "does not support the operation with the supplied arguments", for whatever reason. E.g. ints/datetimes do that:

In [32]: d = datetime.datetime.now()

In [33]: x = 1

In [34]: d.__add__(x)
Out[34]: NotImplemented

In [35]: x.__add__(d)
Out[35]: NotImplemented

In [36]: d + x
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-36-aa9ad8fe26a5> in <module>()
----> 1 d + x

TypeError: unsupported operand type(s) for +: 'datetime.datetime' and 'int'

Surprisingly, though, strings do raise a TypeError. BUT they also call __radd__ if it's present on the other object:

In [37]: 'foobar'.__add__(x)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-37-6c317551e6f0> in <module>()
----> 1 'foobar'.__add__(x)

TypeError: cannot concatenate 'str' and 'int' objects

In [38]: class Foo(object):
   ....:     def __radd__(self, other):
   ....:         return 'xyzzy'
   ....:     

In [39]: 'foobar'.__add__(Foo())
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-39-d607980902de> in <module>()
----> 1 'foobar'.__add__(Foo())

TypeError: cannot concatenate 'str' and 'Foo' objects

In [40]: 'foobar' + Foo()
Out[40]: 'xyzzy'

[jreback]

I guess i disagree with this spec a bit. pandas just tries to do the right thing and not raise too many TypeErrors/NotImplemented.