API: consistency in .loc indexing when no values are found in a list-like indexer GH7999)

[jreback]

closes #7999

[jreback]

@jorisvandenbossche this also fixes the case we talked about (forgot where), in a multiindex when no values are found it will now raise KeyError (I think it would return the original previously)

[jorisvandenbossche]

@jreback Did you also fix the inconsistency between df.loc[[1, 3]] (works as one label is in the index) and df.loc[[1, 3],:] (fails with KeyError)? (as I don't see it mentioned in the whatsnew entry, and they should both work)

Secondly, you now solve the inconsistency between df.loc[[3]] and df.loc[[3], :] in favor of the behaviour of df.loc[[3], :]. But is there a good reason? We could also opt for the behaviour of the single axis (df.loc[[3]]), which is maybe(?) more used ànd is in line with reindex (+ see above, this is rather a bug in the df.loc[list, :] case as it definitely fails with if there is at least 1 label found in the index)

[jreback]

yes, this fixed df.loc[[1,3]] and df.loc[[1,3],:] issue as well (it was the same issue):

In [1]: df = DataFrame([['a'],['b']],index=[1,2])

In [2]: df
Out[2]:
0
1 a
2 b

In [3]: df.loc[[1, 3]]
Out[3]: 
     0
1    a
3  NaN

In [4]: df.loc[[1, 3],:]
Out[4]: 
     0
1    a
3  NaN

The reason I think that the indexer works liek this is: say you have a big list of labels you are getting. It is better to have it work then raise if 1 is off (and just reindex).

Conversely if nothing is found user prob screwed up so KeyError.

This PR doesn't really change much, just makes things consistent (wether you specify all axes or not).

[jorisvandenbossche]

@jreback yes, that makes sense. The only 'pity' is that it is a deviation from reindex, because if you say "df.loc with a list is just like reindex" you have to add "apart from the case if is of length 1 ..."

For df.loc[[1,3],:] do you add that do the whatsnew entry?

[jreback]

cc @ruidc

going to do a do modification? (to include the un-documented prior behavior)

@jorisvandenbossche yeah I know this is different than reindex, but not sure what to do about it.

[jorisvandenbossche]

So, for the whatsnew: just something like (after the other one): There was also a difference between ``df.loc[[1,3]]`` (returns a frame reindexed by [1, 3]) and ``df.loc[[1, 3],:]``(would raise ``KeyError`` prior to 0.15.0). Both will now return a reindexed frame.

[jreback]

ok, sure

[jreback]

ok, updated

[ruidc]

just a typo: line 1254 of indexing.py:
raise KeyError("Not of [%s] are in the [%s]" %
should be "Not ALL of [%s]"...