Broadcasting bug with datetime.date index?

[mcobzarenco]

Guys,

It seems to be a problem with broadcasting when having a datetime.date index, do you know what may be causing this?

Example:

index = index=[datetime.date(2012, 1, 1), datetime.date(2012, 1, 2)]
columns=['a', 'b']
df = pandas.DataFrame([[1., 2.], [3., 4.]], index, columns)

>>> print( df.mean(1) )
2012-01-01    1.5
2012-01-02    3.5

>>> print( (df - df.mean(1)).to_string() )
             a   b  2012-01-01  2012-01-02
2012-01-01 NaN NaN         NaN         NaN
2012-01-02 NaN NaN         NaN         NaN

The columns somehow end up being the union of the inital columns and the index. The above code works OK when using datetime.datetime instead.

Many thanks,
Marius

[changhiskhan]

Try df.sub(df.mean(1), axis=0) as a workaround.

By default it's suppose to broadcast over the columns, except in the case of datetimes where it broadcasts over the index by default because it's such a common operation. I think we should expand it to cover dates and times as well, @wesm what do you think? Purist in me says maybe get rid of this hack altogether, but the use case is compelling

[wesm]

Binary ops between DataFrame of time series and a single time series has been special-cased to go column-by-column. This is some serious legacy stuff from pandas 0.1 and should probably receive a deprecation message in 0.10, to get users to change their code to use one of the explicit functions

[changhiskhan]

Wow, I didn't know it went THAT far back...not sure how much legacy code is built on that assumption. Probably going to make a certain group of devs at a certain company fairly unhappy.