API: Change MultiIndex.levels/codes to use `tuple` instead of `FrozenList`?

[mroeschke]

The doc of FrozenList says

    Container that doesn't allow setting item *but*
    because it's technically hashable, will be used
    for lookups, appropriately, etc.

which a tuple satisfies. Not sure why a tuple wasn't use originally but it may be simpler to use them instead of a custom FrozenList that acts like a tuple.

[srkds]

Hi,

While reading about Frozenlist I found a StackOverflow conversation (Link) and an answer from @jreback regarding the same.

The point of it is to prevent modification of these thru attributes and force the use of methods (e.g. set_levels()).

And If planning to change to tuple is it the correct way?

- def levels(self) -> FrozenList:
+ def levels(self) -> tuple:
        # Use cache_readonly to ensure that self.get_locs doesn't repeatedly
        # create new IndexEngine
        # https://github.com/pandas-dev/pandas/issues/31648
        result = [x._rename(name=name) for x, name in zip(self._levels, self._names)]
        for level in result:
            # disallow midx.levels[0].name = "foo"
            level._no_setting_name = True
-        return FrozenList(result)
+        return tuple(result)

expected output

>>> import pandas as pd
>>> import numpy as np
>>> arrays = [
...     np.array(["bar", "bar", "baz", "baz", "foo", "foo", "qux", "qux"]),
...     np.array(["one", "two", "one", "two", "one", "two", "one", "two"]),
... ]
>>> tuples = list(zip(*arrays))
>>> index = pd.MultiIndex.from_tuples(tuples, names=["first", "second"])
>>> df = pd.DataFrame(np.random.randn(3, 8), index=["A", "B", "C"], columns=index)
>>> df.columns.levels
(Index(['bar', 'baz', 'foo', 'qux'], dtype='object', name='first'), Index(['one', 'two'], dtype='object', name='second'))
>>>

[mroeschke]

Nice find @srkds!

I'm not sure if any of those reason are still entirely applicable at this point. I would say if the custom FrozenList methods are not used extensively (e.g. union), it might as well just return a tuple as it would be better that the returned object is a public class (FrozenList is only accessible from the private pandas.core)

And If planning to change to tuple is it the correct way?

Yeah exactly

[srkds]

I'm not sure if any of those reason are still entirely applicable at this point. I would say if the custom FrozenList methods are not used extensively (e.g. union), it might as well just return a tuple as it would be better that the returned object is a public class (FrozenList is only accessible from the private pandas.core)

Yeah sounds good.

Yeah exactly

So going for submitting a PR.

[srkds]

take

[mroeschke]

@srkds I would wait until there's a little more feedback before submitting a PR. Maybe there's a reason someone else might bring up why we should still use FrozenList

[srkds]

@srkds I would wait until there's a little more feedback before submitting a PR. Maybe there's a reason someone else might bring up why we should still use FrozenList

Yes, I agree with you👍

[topper-123]

I think it can be done for levels/codes. So I'm +1, if there's no surprises after removing FrozenList.

We still will use FrozenList in .names:

>>> mi = pd.MultiIndex.from_frame(df, names=[(1,), (4,)])
>>> mi.names
FrozenList([(1,), (4,)])

[mroeschke]

Looks pretty feasible internally based on #53582, but I think this may need to a hard 3.0 break as I don't think there's a clean way to deprecate this

[topper-123]

Yeah, agree with that.