ENH: drop_level argument for xs

[hayd]

As discussed here, xs usually remove the level which you are accessing. This commit allows you to explicitly say you don't want that (and want to keep the same levels in the result).

In [4]: df = DataFrame({'day': {0: 'sat', 1: 'sun'},
                                   'flavour': {0: 'strawberry', 1: 'strawberry'},
                                   'sales': {0: 10, 1: 12},
                                   'year': {0: 2008, 1: 2008}}).set_index(['year','flavour','day'])

In [5]: df
Out[5]:
                     sales
year flavour    day
2008 strawberry sat     10
                sun     12

In [6]: df.xs('sat', level='day')
Out[6]:
                 sales
year flavour
2008 strawberry     10

In [7]: df.xs('sat', level='day', drop_level=False)
Out[7]:
                     sales
year flavour    day
2008 strawberry sat     10

In [8]: df.xs([2008, 'sat'], level=['year', 'day'], drop_level=False)
Out[8]:
                     sales
year flavour    day
2008 strawberry sat     10

[hayd]

oh dear, screwed something up in this, need to fix.

[cpcloud]

@hayd should be marked for 0.13?

[hayd]

@jreback merge? or do you have a better arg name?

[jreback]

hmmm.....and we don't want to make drop_level a public function....

could we do this thru drop? maybe like this: (same impl)

df.drop('sat', level='day')

?

[hayd]

This makes xs more like a filter (atm the behaviour is to drop the selected level, this adds option not too).

[jreback]

this is adding a modification option xs, something it doesn't do now; that's my hesitation

what about adding drop_level, be very explicit?

[hayd]

ah ok, yeah drop_level is better

[hayd]

Hmmm giving def _drop_levels a drop_level was stupid, this is a bad way to do this. Need to reconsider.

[jreback]

no...I meant in core/generic.py make a method called drop_levels with same args as xs, label, axis, level and just do what you are doing now

[hayd]

ah, true and this isn't really a cross section either... I dunno actually I think this makes sense as an option for xs...just viewing the data in slightly different way, with superfluous index level.

I guess you can hack select to do this:

In [10]: df.select(lambda x: x[2] == 'sat')
Out[10]:
                     sales
year flavour    day
2008 strawberry sat     10

Maybe could just add this api to select? (add level argument)

df.select(self, crit, axis=0, level=None)

looks to see if crit is callable, if is current behaviour, if not then looks at level whatever and does xs without dropping level (the result we're talking about).

[jreback]

would shy away from select, its a linear search...let's go back to xs...

how about ''drop=...'' ? (rather than drop_selected_level?

[hayd]

I liked you drop_level (as a compromise :p), I think it's clearer what you're dropping then. Ok, have to clean up this code though, it looks like I just threw this together without thinking about back-end semantics. I think it's much easier than my awful hack.

[hayd]

@jreback logic is now much improved.

[jreback]

looks good..