API: should .loc be as strict as .ix in tuple indexing?

[jreback]

from ML

https://groups.google.com/forum/?fromgroups=#!topic/pydata/fzPGjVrj9lU

df = pd.DataFrame.from_items([('a', [1,2]), ('b', ([3,4]))])
df.index = pd.MultiIndex.from_tuples([('A', 'B'),('C', 'D')],
        names=['k1', 'k2'])

# A dataframe, with a multiindex:
>>> df
       a  b
k1 k2
A  B   1  3
C  D   2  4

.ix[('A','B')] succeeds
.ix['A','B'] succeeds as the A, B is tried as a tuple (.ix tries hard here)

propose that we change .loc to disallow the 2nd case and instead raise a KeyError
its more 'strict'

[ghost]

IIRC, that's a python wart in __getitem__, those two cases look identical
to the callee:

In [2]: class A(object):
   ...:     def __getitem__(self,*args):
   ...:         print args
   ...: a=A()

In [3]: a['1','2']
(('1', '2'),)

In [4]: a[('1','2')]
(('1', '2'),)

[jreback]

@y-p ahh yes.........I guess there is not a way to disambiguate if the tuple exists in a mi in the index, versus the case where the individual elements existing in both an index and a column......

[ghost]

yeah. I think in some places in pandas the type is used to dismbiguate behaviour
(tuple vs. list trigger different code paths). I noticed a while back that the semantics
of sets aren't exploited yet. Maybe we can put that to good use someday
(with slighly verbose syntax on python 2.6), as if pandas' signature overloading wasn't
ridiculous already

What should .iloc[{1,2,3,5}] mean? random permutation of rows?
label-uniquified slice? foo? bar?

[jreback]

:)....going to push to 0.12...to see if we want to change the API at all

[jreback]

closing as impossible to disambiguate w/o new syntax which makes things even more magical than they are now