Using .loc.__setitem__ with integer slices does not raise

[lumbric]

Code Sample

>>> d = pd.DataFrame({'a': range(5), 'b': range(2,7), 'c': range(3,8)})
>>> d
   a  b  c
0  0  2  3
1  1  3  4
2  2  4  5
3  3  5  6
4  4  6  7
>>> d.loc[0:2, 0:2] = 42   # note that loc is used with an integer slice! should this raise?
>>> d  # this is somehow the expected result
    a   b  c
0  42  42  3
1  42  42  4
2  42  42  5
3   3   5  6
4   4   6  7
>>> d.iloc[0:2, 0:2]        # expected result
    a   b
0  42  42
1  42  42
>>> d.loc[0:2, 0:2]        # this is the same syntax as above, but raises
...
TypeError: cannot do slice indexing on <class 'pandas.core.indexes.base.Index'> with these indexers [0] of <class 'int'>

Problem description

I'd expect a symmetry in d.loc.__getitem__ and d.loc.__setitem__. At the moment only getitem raises if integer slices are passed. Usage of .loc was probably a mistake, instead .iloc should have been used. Even if it works correctly, it is confusing for users.

Probably this is simply a missing exception in the special case of d.loc.__setitem__ being called with integer slices.

See #13831 for a very similar problems with pd.Series.

Expected Output

Assuming d.loc.__getitem__ should raise for integer slices (as it does), then also d.loc.__setitem__ should raise:

>>> d.loc[0:2, 0:2]
...
SomException: int & loc don't go together

Output of pd.show_versions()

INSTALLED VERSIONS ------------------ commit: None python: 3.6.8.final.0 python-bits: 64 OS: Linux OS-release: 4.18.0-20-generic machine: x86_64 processor: x86_64 byteorder: little LC_ALL: None LANG: en_US.UTF-8 LOCALE: en_US.UTF-8

pandas: 0.24.2
pytest: 3.6.4
pip: 9.0.1
setuptools: 40.8.0
Cython: None
numpy: 1.16.2
scipy: 1.2.1
pyarrow: None
xarray: 0.11.3
IPython: 5.5.0
sphinx: 1.7.9
patsy: 0.4.1+dev
dateutil: 2.8.0
pytz: 2019.1
blosc: 1.5.1
bottleneck: 1.2.1
tables: 3.4.4
numexpr: 2.6.5
feather: None
matplotlib: 2.2.2
openpyxl: None
xlrd: None
xlwt: None
xlsxwriter: None
lxml.etree: 4.2.5
bs4: 4.5.0
html5lib: 0.999999999
sqlalchemy: None
pymysql: None
psycopg2: 2.7.5 (dt dec pq3 ext lo64)
jinja2: 2.10
s3fs: None
fastparquet: None
pandas_gbq: None
pandas_datareader: None
gcsfs: None

[jorisvandenbossche]

@lumbric Thanks a lot for the report. That's certainly a bug as there should indeed be a symmetry between setitem/getitem (although not always, you can set with new labels to expand a dataframe in certain cases), and in any case: it is interpreting the 0:2 as positional which is the guarantee of loc that this should not happen.

It only seems to be the case for slicing, as a list of integers correctly raises:

In [9]: d.loc[0:2, [0, 1]] = 42
...
KeyError: "None of [Int64Index([0, 1], dtype='int64')] are in the [columns]"

[jorisvandenbossche]

Very welcome to try to look into the loc setitem code path were this may be coming from!