[BUG] .at accessor for new column fails in cudf.pandas

[bdice]

Describe the bug
I was helping a user Amith with a question about cudf.pandas and the library pandapower (see RAPIDS GoAI Slack). The pandapower notebook examples trigger a bug in cudf.pandas around the .at accessor. @Jacey0 helped debug this with me during PyCon 2024. Thanks Jacey!

Steps/Code to reproduce bug

Here is a Colab notebook with a reproducer: https://colab.research.google.com/gist/bdice/f5be320dada30671015d5b53642ec19c/solution-test.ipynb

Here is the minimal code snippet needed to reproduce:

%load_ext cudf.pandas
import pandas as pd

bus = pd.DataFrame({'name': []})
entries = {'name': 'Bus 0', 'geo': 'X and Y'}
for col, val in entries.items():
    print("Setting", col, "to", val, end="\n\n")
    print("Before:")
    print(bus, end="\n\n")
    bus.at[0, col] = val # pandas adds a new row and 'geo' column' fine
    print("After:")
    print(bus, end="\n\n")

Error output:

Setting name to Bus 0

Before:
Empty DataFrame
Columns: [name]
Index: []

---------------------------------------------------------------------------

KeyError                                  Traceback (most recent call last)

/usr/local/lib/python3.10/dist-packages/cudf/core/series.py in _loc_to_iloc(self, arg)
    381                 if (n := len(indices)) == 0:
--> 382                     raise KeyError("Label scalar is out of bounds")
    383                 elif n == 1:

KeyError: 'Label scalar is out of bounds'


During handling of the above exception, another exception occurred:

KeyError                                  Traceback (most recent call last)

15 frames

KeyError: 'Label scalar is out of bounds'


During handling of the above exception, another exception occurred:

IndexError                                Traceback (most recent call last)

IndexError: Index out of bounds


During handling of the above exception, another exception occurred:

AttributeError                            Traceback (most recent call last)

/usr/local/lib/python3.10/dist-packages/pandas/core/indexing.py in __setitem__(self, key, value)
    843         else:
    844             key = com.apply_if_callable(key, self.obj)
--> 845         indexer = self._get_setitem_indexer(key)
    846         self._has_valid_setitem_indexer(key)
    847 

AttributeError: '_AtIndexer' object has no attribute '_get_setitem_indexer'

Expected behavior
Here is the output from plain pandas:

Setting name to Bus 0

Before:
Empty DataFrame
Columns: [name]
Index: []

After:
    name
0  Bus 0

Setting geo to X and Y

Before:
    name
0  Bus 0

After:
    name      geo
0  Bus 0  X and Y

[mroeschke]

It appears the cudf.pandas bug is that _AtIndexer (and _iAtIndexer) indexer are not defined in _wrappers/pandas.py as proxy objects.

And of course, it appears cudf.DataFrame.loc.__setitem__ that expands is also buggy

[mroeschke]

More investigation notes:

So cudf.DataFrame.at is just returns cudf.DataFrame.loc, so a pandas.DataFrame.at call will always succeed to return cudf.pandas' intermediate proxy of pandas.DataFrame.loc.

So if pandas.DataFrame.at.__setitem__ fails in cudf.pandas, we rewind to pandas.DataFrame.loc.__setitem__ instead of pandas.DataFrame.at.__setitem__.

I think we need pandas.DataFrame.at to be an IntermediateProxy of an IntermediateProxy as "proper" fix. I'm not sure if there's an straightforward way to do this today.

Alternatively, if we make cudf.DataFrame.at not map to cudf.DataFrame.loc (i.e. return a different object with the same implementation), that would be an easier fix...

[bdice]

Alternatively, if we make cudf.DataFrame.at not map to cudf.DataFrame.loc (i.e. return a different object with the same implementation), that would be an easier fix...

That sounds like the right thing to do. cudf's object model should mirror that of pandas. If at returns loc in cudf but not pandas, that's a problem in the mirrored object model.