API: Use different exception type when MultiIndex indexing fails due to index not being lexsorted

[Dr-Irv]

I'd like to be able to separately trap errors in indexing that are due to the lack of the MultiIndex being lexsorted. Right now, a KeyError is raised, and the only way to tell if the error was due to the lack of a lexsort versus a missing key is to look at the text of the message.

So I'd like to avoid code like this:

try:
    subdf = df.loc['A':'D','Vals']
except KeyError as ker:
    if ker.args[0].startswith("MultiIndex Slicing requires the index to be fully lexsorted"):
        print ("Need to handle fact index not sorted")
    else:
        print ("Need to handle fact that key was missing")

I'd rather write something like:

try:
    subdf = df.loc['A':'D','Vals']
except UnsortedIndexError:
    print ("Need to handle fact index not sorted")
except KeyError:
    print ("Need to handle fact that key was missing")

So could the designers accept a change where a different error is raised in case the index is not lexsorted? If so, I'll implement it.

[jreback]

pls show a complete example uptop

[Dr-Irv]

Here is an example:

mi = pd.MultiIndex.from_tuples([('z','a'),('x','a'),('y','b'),('x','b'),('y','a'),('z','b')], names=['one','two'])
df = pd.DataFrame([[i,10*i] for i in range(6)], index=mi, columns=['one','two'])
print(df.loc(axis=0)['z',:])

The last line fails with:
KeyError: 'MultiIndex Slicing requires the index to be fully lexsorted tuple len (2), lexsort depth (0)'

Now consider where we do the sorting:

df.sort_index(inplace=True)
print(df.loc(axis=0)['z',:])
print(df.loc(axis=0)['q',:])

The second line works because things are now sorted, but the last line fails with:
KeyError: 'q'

So I'd like to be able to surround all of the above code with ways of detecting the 2 different errors, which will make debugging easier.

[shoyer]

I like this idea. To preserve backwards compatibility, we might make it a KeyError subclass. Note that there are also some cases where indexing non-multiindexes can raise a KeyError because the index is unsourced (notably, when slicing).

[jreback]

would for sure have to be a KeyError for back-compat. (ideally this should be a ValueError, but that boat has sailed).

[Dr-Irv]

@jreback So should I create the UnsortedIndexError code, test it, and do the pull request?

[jreback]

@Dr-Irv interested in working on this?

[Dr-Irv]

Yes, but it might be a while, as I've got some other pressing issues.