REGR: GH50127 fixes error for converting to / creating Index with dtype S3

doc/source/whatsnew/v2.0.0.rst

@@ -1399,6 +1399,7 @@ Other
 - Bug in incorrectly accepting dtype strings containing "[pyarrow]" more than once (:issue:`51548`)
 - Bug in :meth:`Series.searchsorted` inconsistent behavior when accepting :class:`DataFrame` as parameter ``value`` (:issue:`49620`)
 - Bug in :func:`array` failing to raise on :class:`DataFrame` inputs (:issue:`51167`)
+- Bug in :class:`Index` when created or converted to the numpy dtype S3 would raise ``NotImplementedError`` (:issue:`50127`)
 
 .. ---------------------------------------------------------------------------
 .. _whatsnew_200.contributors:

pandas/core/indexes/base.py

@@ -512,6 +512,11 @@ def __new__(
  # they are actually ints, e.g. '0' and 0.0
  # should not be coerced
  data = com.asarray_tuplesafe(data, dtype=_dtype_obj)
+ # GH#50127 we must update the `dtype` when we have the numpy
+ # type `S` to `_dtype_to_subclass`, because it would raise a
+ # `NotImplementedError`.
+ if dtype and dtype.kind == "S":
+ dtype = _dtype_obj
 
  elif is_scalar(data):
  raise cls._raise_scalar_data_error(data)
@@ -543,10 +548,13 @@ def __new__(
  if len(data) == 0:
  # unlike Series, we default to object dtype:
  data = np.array(data, dtype=object)
-
- if len(data) and isinstance(data[0], tuple):
+ elif isinstance(data[0], tuple):
  # Ensure we get 1-D array of tuples instead of 2D array.
  data = com.asarray_tuplesafe(data, dtype=_dtype_obj)
+ elif dtype and dtype.kind == "S":
+ # GH#50127 we update data to a np.array with the correct dtype.
+ data = astype_array(np.array(data, copy=copy), dtype=dtype, copy=copy)
+ dtype = _dtype_obj
 
  try:
  arr = sanitize_array(data, None, dtype=dtype, copy=copy)

pandas/tests/dtypes/test_dtypes.py

@@ -18,6 +18,7 @@
  is_interval_dtype,
  is_period_dtype,
  is_string_dtype,
+ pandas_dtype,
 )
 from pandas.core.dtypes.dtypes import (
  CategoricalDtype,
@@ -1122,3 +1123,15 @@ def test_multi_column_dtype_assignment():
 
  df["b"] = 0
  tm.assert_frame_equal(df, expected)
+
+
+def test_numpy_s3_dtype_on_index():
+ # GH #50127
+ index = pd.Index(["abcd", "1234"], dtype="S3")
+ expected = pd.Index(["abc", "123"], dtype="S3")
+ assert index.dtype == pandas_dtype("object")
+ tm.assert_index_equal(index, expected)
+
+ index = pd.Index(["abcd", "1234"])
+ expected = pd.Index(["abc", "123"], dtype="S3")
+ tm.assert_index_equal(index.astype("S3"), expected)

pandas/tests/io/test_parquet.py

@@ -1036,9 +1036,7 @@ def test_columns_dtypes_not_invalid(self, pa):
 
  # bytes
  df.columns = [b"foo", b"bar"]
- with pytest.raises(NotImplementedError, match="|S3"):
- # Bytes fails on read_parquet
- check_round_trip(df, pa)
+ check_round_trip(df, pa)
 
  # python object
  df.columns = [