DataFrame.iloc does not raise for setitem if empty

[SaturnFromTitan]

This is a follow-up on #32886 (comment)

test_setitem_ndarray_3d in pandas/tests/indexing/test_indexing.py isn't raising for empty index, iloc and DataFrame parametrization.

The mentioned PR added a new xfail test for this combination: test_setitem_ndarray_3d_does_not_fail_for_iloc_empty_dataframe. The goal of this PR is to remove the xfail test and remove the associated pytest.skip in test_setitem_ndarray_3d.

[simonjayhawkins]

This didn't silently fail on 0.25.3

>>> import numpy as np
>>>
>>> import pandas as pd
>>>
>>> pd.__version__
'0.25.3'
>>>
>>> i = pd.Index([])
>>> obj = pd.DataFrame(np.random.randn(len(i), len(i)), index=i, columns=i)
>>> obj
Empty DataFrame
Columns: []
Index: []
>>>
>>> nd3 = np.random.randint(5, size=(2, 2, 2))
>>>
>>> obj.iloc[nd3] = 0
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Users\simon\Anaconda3\lib\site-packages\pandas\core\indexing.py", line 204, in __setitem__
    indexer = self._get_setitem_indexer(key)
  File "C:\Users\simon\Anaconda3\lib\site-packages\pandas\core\indexing.py", line 191, in _get_setitem_indexer
    return self._convert_to_indexer(key, axis=axis, is_setter=True)
  File "C:\Users\simon\Anaconda3\lib\site-packages\pandas\core\indexing.py", line 2175, in _convert_to_indexer
    self._validate_key(obj, axis)
  File "C:\Users\simon\Anaconda3\lib\site-packages\pandas\core\indexing.py", line 2031, in _validate_key
    raise IndexError("positional indexers are out-of-bounds")
IndexError: positional indexers are out-of-bounds

922f932 is the first bad commit
commit 922f932
Author: jbrockmendel [email protected]
Date: Thu Feb 27 04:52:36 2020 -0800

PERF: pass through to numpy validation for iloc setitem (#32257)

cc @jbrockmendel

[jbrockmendel]

This matches the numpy behavior:

arr = np.array([]).reshape(0, 0)
nd3 = np.random.randint(5, size=(2, 2, 2))
arr[nd3] = 0

This may be a numpy bug as well cc @seberg ?

[seberg]

Yeah, I guess it technically is a NumPy bug. The reason here must be the additional empty dimension that is not being indexed, somehow causing the index check to never be run to begin with.
Can you open a NumPy issue about it? I am surprised there is a regression, this is probably the case at least since NumPy 1.9, but then I guess the regression is due to a change in how pandas does things internally?

[jbrockmendel]

but then I guess the regression is due to a change in how pandas does things internally?

yes, we recently stopped doing some validation on iloc and instead fall through to let numpy do the validation

[seberg]

As a heads up, I am considering deprecating this for a release if that is OK? I agree that it is a straight bug, but apparently I intentionally put this in (although possibly not intentionally this exact case).
If it would be easier for you if I remove it right away, that would be a welcome vote though :) (may still decide to do that in any case).

[jbrockmendel]

So we'll put back in a check and raise, and then when our min numpy version catches up to 1.19(?) we can go back to falling through? If so, sounds good.

[seberg]

Yeah, I suppose so. You could get away with only doing it when the result/assignment may be empty (if you want to avoid doing unnecessary/slow checks). If we deprecate, I guess it would be 1.20 otherwise 1.19, but since you need backcompat anyway, I guess it makes no real difference.

[phofl]

cc @jbrockmendel This raises since cd649a2

Traceback (most recent call last):
  File "/home/developer/.config/JetBrains/PyCharm2020.2/scratches/scratch_4.py", line 145, in <module>
    obj.iloc[nd3] = 0
  File "/home/developer/PycharmProjects/pandas/pandas/core/indexing.py", line 687, in __setitem__
    iloc._setitem_with_indexer(indexer, value, self.name)
  File "/home/developer/PycharmProjects/pandas/pandas/core/indexing.py", line 1630, in _setitem_with_indexer
    self._setitem_with_indexer_split_path(indexer, value, name)
  File "/home/developer/PycharmProjects/pandas/pandas/core/indexing.py", line 1646, in _setitem_with_indexer_split_path
    raise ValueError(r"Cannot set values with ndim > 2")
ValueError: Cannot set values with ndim > 2

Can we close this ?