BUG: raise on empty-string conversion to pyarrow float (#66834)

Empty CSV fields with keep_default_na=False stored IEEE NaN in double[pyarrow] instead of raising like numpy float64.

Author: coderunlouis

doc/source/whatsnew/v3.1.0.rst

@@ -649,7 +649,7 @@ Missing
 - Bug in :meth:`DataFrame.fillna` with a dict value raising ``RecursionError`` when columns are a :class:`MultiIndex` with duplicate entries (:issue:`53498`)
 - Bug in :meth:`DataFrame.interpolate` and :meth:`Series.interpolate` with ``method`` in ``"index"``, ``"values"`` or ``"time"`` raising when the index had an :class:`ArrowDtype` timestamp or duration dtype; these now match the equivalent :class:`DatetimeIndex` or :class:`TimedeltaIndex` (:issue:`66338`)
 - Bug in :meth:`Series.combine_first` crashing when Series names are :class:`Timestamp` objects (:issue:`65333`)
-- Bug in :class:`ArrowExtensionArray` where empty strings converted to a floating PyArrow dtype stored IEEE ``NaN`` instead of null, so :meth:`Series.fillna` did not fill them (e.g. empty CSV fields with ``dtype="double[pyarrow]"`` and ``keep_default_na=False``)
+- Bug in :class:`ArrowExtensionArray` where converting empty strings to a floating PyArrow dtype stored IEEE ``NaN`` instead of raising like numpy ``float64`` (e.g. empty CSV fields with ``dtype="double[pyarrow]"`` and ``keep_default_na=False``) (:issue:`66834`)
 
 MultiIndex
 ^^^^^^^^^^

pandas/core/arrays/arrow/array.py

@@ -514,12 +514,12 @@ def _from_sequence_of_strings(
                 scalars = strings.cast(pa_type)
             else:
                 mask = isna(strings)
-                if (
-                    is_nan_na()
-                    and isinstance(scalars, np.ndarray)
-                    and np.issubdtype(scalars.dtype, np.floating)
-                ):
-                    mask = np.asarray(mask, dtype=np.bool_) | np.isnan(scalars)
+                # to_numeric("") yields IEEE NaN without raising. Match numpy
+                # float64: an empty string that was not treated as NA is invalid.
+                arr = np.asarray(strings, dtype=object)
+                unmasked_empty = (arr == "") & ~np.asarray(mask, dtype=np.bool_)
+                if unmasked_empty.any():
+                    raise ValueError("could not convert string to float: ''")
                 if mask is not None:
                     scalars = pa.array(scalars, mask=mask, type=pa_type)
 

pandas/tests/extension/test_arrow.py

@@ -3562,18 +3562,20 @@ def test_from_sequence_of_strings_boolean():
         ArrowExtensionArray._from_sequence_of_strings(strings, dtype=dtype)
 
 
-def test_from_sequence_of_strings_empty_string_float(using_nan_is_na):
+def test_from_sequence_of_strings_empty_string_float():
+    # GH#66834 match numpy float64: empty string is not a valid float
     strings = ["1.5", "", "2.0"]
     dtype = ArrowDtype(pa.float64())
+    with pytest.raises(ValueError, match="could not convert string to float"):
+        ArrowExtensionArray._from_sequence_of_strings(strings, dtype=dtype)
+
+
+def test_from_sequence_of_strings_none_float():
+    strings = ["1.5", None, "2.0"]
+    dtype = ArrowDtype(pa.float64())
     result = ArrowExtensionArray._from_sequence_of_strings(strings, dtype=dtype)
-    if using_nan_is_na:
-        expected = ArrowExtensionArray(pa.array([1.5, None, 2.0], type=pa.float64()))
-        tm.assert_extension_array_equal(result, expected)
-        filled = pd.Series(result, dtype=dtype).fillna(0)
-        tm.assert_series_equal(filled, pd.Series([1.5, 0.0, 2.0], dtype=dtype))
-    else:
-        assert not result.isna().any()
-        assert np.isnan(result.to_numpy(dtype="float64")[1])
+    expected = ArrowExtensionArray(pa.array([1.5, None, 2.0], type=pa.float64()))
+    tm.assert_extension_array_equal(result, expected)
 
 
 def test_concat_empty_arrow_backed_series(dtype):