expand comments

Signed-off-by: Anatoly Myachev <[email protected]>

modin/core/dataframe/pandas/dataframe/dataframe.py

@@ -3985,6 +3985,7 @@ def _compute_new_widths():
  new_columns = joined_index
  frames = [self] + others
  # TODO: should we wrap all `concat` call into "try except" block?
+ # `ModinDtypes.concat` can throw exception in case of duplicate values
  new_dtypes = ModinDtypes.concat([frame._dtypes for frame in frames], axis=1)
  # If we have already cached the length of each row in at least one
  # of the row's partitions, we can build new_lengths for the new
@@ -4621,6 +4622,9 @@ def _arrow_type_to_dtype(cls, arrow_type):
 
  try:
  # TODO: should we map arrow types to pyarrow-backed pandas types?
+ # It seems like this might help avoid the expense of transferring
+ # data between backends (numpy and pyarrow), but we need to be sure
+ # how this fits into the type inference system in pandas.
  res = arrow_type.to_pandas_dtype()
  # Conversion to pandas is not implemented for some arrow types,
  # perform manual conversion for them:

modin/core/storage_formats/pandas/groupby.py

@@ -360,6 +360,7 @@ def applyier(df, other): # pragma: no cover
  # different partitions
  if len(index) == 0 and len(columns) > 0:
  common_type = find_common_type(result.dtypes.tolist())
+ # TODO: remove find_common_type+astype after pandas fix the following issue
  # transpose loses dtypes: https://github.com/pandas-dev/pandas/issues/43337
  result = result.transpose().astype(common_type, copy=False)
 

modin/tests/pandas/dataframe/test_binary.py

@@ -87,7 +87,7 @@ def test_math_functions(other, axis, op, backend):
  pytest.xfail(reason="different behavior")
 
  if op in ("mod", "rmod") and backend == "pyarrow":
- pytest.skip(reason="Not implemented for pyarrow backend")
+ pytest.skip(reason="These functions are not implemented in pandas itself")
  eval_general(
  *create_test_dfs(data, backend=backend),
  lambda df: getattr(df, op)(other(df, axis), axis=axis),