Dropping top level in multiindex independently of key

[didacrarbones]

I am developing some code that derives various pandas classes, and I have stumbled what seems to me like a bug. I do not have too much experience with pandas, so pardon me if this is obviously wrong.

In function maybe_droplevels from the multiindex file, lines 3878 and 3884 disregard the content of key and directly drop the top-most level. It has been like that for at least some years. I am not sure if this was by design in the beginning, to avoid issues, but after a brief look I cannot find the reason why it is would not possible to drop only the level where key is.

My code derives MultiIndex, and when I return a slice that shrinks the DataFrame size, line 3463 wants to shrink the size of the index too.

[phofl]

Hi, could you please provide something reproducible showing the bug?

[attack68]

the function

def maybe_droplevels(index: Index, key) -> Index:
    # drop levels
    original_index = index
    if isinstance(key, tuple):
        for _ in key:
            try:
                index = index._drop_level_numbers([0])
            except ValueError:
                # we have dropped too much, so back out
                return original_index
    else:
        try:
            index = index._drop_level_numbers([0])
        except ValueError:
            pass

    return index

does indeed only drop the top levels, effectively it is doing the same as (which I find more explicit):

def maybe_droplevels(index: Index, key) -> Index:
    # drop leading levels
    if isinstance(key, tuple):
        if len(key) >= index.nlevels: # cannot drop all or more
            return index
        else:  # drop the leading number corresponding to key length
            return index._drop_level_numbers(list(range(len(key)))        
    else:
        if index.nlevels == 1:
            return index
        else:  # drop leading level
            return index._drop_level_numbers([0])

Sorry, without further investigation I too don't know why this would be the case.

I might speculate the when using .loc with a multiIndex you do not have to index with a a tuple the length of nlevels but if you want to index say the 2nd or 3rd levels, you must always populate the left-most, i.e. .loc[(slice(None), "a", "z")]`, this might be relevant for this method. But this may also be a red-herring so beware!

[mroeschke]

Thanks for the report. While this does point to suspicious behavior, it would be great to have a small reproducible example of the discrepancy so we could ultimately see the difference and have a unit test as well. Closing, but happy to reopen once we can get that information.