Index API proposal: unified axis label lookup

[immerrr]

As the next step of separation-of-concerns plan (#6744) I'd like to
propose adding a method (or several, actually) to Index class that
would encapsulate the details of foo.loc[l1,l2,...] lookup.

Implementation Idea

Roughly, the idea is to make loc's getitem as simple as

def __getitem__(self, indexer):
    axes = self.obj.axes
    return self.obj.iloc[axes[0].lookup_labels_nd(indexer, axes[1:], typ='loc')]

Not quite, but hopefully you get the point. The default lookup_labels_nd implementation would then look something like this:

def lookup(self, indexer, other_axes, typ=None):
    if not isinstance(indexer, tuple):
        return self.lookup_labels(indexer, typ=typ)
    else:
        # ndim mismatch error handling is omitted intentionally
        return (self.lookup_labels(indexer[0]),) + \
               tuple(ax.lookup_labels(ix, typ=typ)
                     for ax, ix in zip(other_axes, indexer))

The result should be an object that could be fed to an underlying
BlockManager to perform the requested operation. To support adding
new rows with "setitem", it is only needed to agree that lookup_labels_nd will
never return negative indices unless they reference newly appended
items along that axis.

This would allow to hide Index-subclass-specific lookup peculiarities
in their respective overrides of lookup_labels_nd and lookup_labels (proposals for
better names are welcome), e.g.:

• looking up str in DatetimeIndex/PeriodIndex
• looking up int in FloatIndex
• looking up per-level slices in MultiIndex

Benefits

• no more confusing errors due to try .. catch block carpet-catching a
  logic error, because corner cases will be handled precisely where
  they are needed and nowhere else
• no more relying on isinstance checks and exceptions to seek for
  alternative lookup scenarios, meaning more performance
• the API will provide a contract that is simple to grasp, test, benchmark and,
  eventually, cythonize (as a side effect of this point I'd like to try putting
  up a wiki page with indexing API reference)

[immerrr]

I was trying to mock up a proof of concept (while at ep2014) and decided to start with simple things, at and iat, but found something weird:

In [19]: s = pd.Series([1,2,3], index=list('abc'))

In [20]: s.at['a']
Out[20]: 1

In [21]: s.at['b']
Out[21]: 2

In [22]: s.at['c']
Out[22]: 3

In [23]: s.at[0]
Out[23]: 1

I couldn't find any reference to this feature in pandas doc, and I have a strong feeling it shouldn't be there. @jrebac, what do you think, do I have to maintain it?

[jreback]

I think that is a bug, at should be a restricted version of loc, e.g. (no slicing or lists), and certainly no fallback. Prob just not checking that the label type is a valid type for the index. pls do a PR to test/fix this if you would. thanks

[jreback]

#7814

[shoyer]

Just stumbled across this -- it gets a bit +1 from me!