Backport PR #52843 on branch 2.0.x (BUG: pyarrow duration arrays constructed from data containing NaT can overflow) (#52869)

BUG: pyarrow duration arrays constructed from data containing NaT can overflow (#52843)

(cherry picked from commit 4539f3e)

Co-authored-by: Luke Manley <[email protected]>

Author: phofl

doc/source/whatsnew/v2.0.1.rst

@@ -28,6 +28,7 @@ Bug fixes
 ~~~~~~~~~
 - Bug in :attr:`Series.dt.days` that would overflow ``int32`` number of days (:issue:`52391`)
 - Bug in :class:`arrays.DatetimeArray` constructor returning an incorrect unit when passed a non-nanosecond numpy datetime array (:issue:`52555`)
+- Bug in :class:`~arrays.ArrowExtensionArray` with duration dtype overflowing when constructed from data containing numpy ``NaT`` (:issue:`52843`)
 - Bug in :func:`Series.dt.round` when passing a ``freq`` of equal or higher resolution compared to the :class:`Series` would raise a ``ZeroDivisionError`` (:issue:`52761`)
 - Bug in :func:`Series.median` with :class:`ArrowDtype` returning an approximate median (:issue:`52679`)
 - Bug in :func:`api.interchange.from_dataframe` was unnecessarily raising on categorical dtypes (:issue:`49889`)

pandas/core/arrays/arrow/array.py

@@ -260,7 +260,13 @@ def _from_sequence(cls, scalars, *, dtype: Dtype | None = None, copy: bool = Fal
                 scalars = pa.array(scalars, from_pandas=True)
         if pa_dtype:
             scalars = scalars.cast(pa_dtype)
-        return cls(scalars)
+        arr = cls(scalars)
+        if pa.types.is_duration(scalars.type) and scalars.null_count > 0:
+            # GH52843: upstream bug for duration types when originally
+            # constructed with data containing numpy NaT.
+            # https://github.com/apache/arrow/issues/35088
+            arr = arr.fillna(arr.dtype.na_value)
+        return arr
 
     @classmethod
     def _from_sequence_of_strings(

pandas/tests/extension/test_arrow.py

@@ -2629,3 +2629,19 @@ def test_describe_numeric_data(pa_type):
         index=["count", "mean", "std", "min", "25%", "50%", "75%", "max"],
     )
     tm.assert_series_equal(result, expected)
+
+
+@pytest.mark.xfail(
+    pa_version_under8p0,
+    reason="Function 'add_checked' has no kernel matching input types",
+    raises=pa.ArrowNotImplementedError,
+)
+def test_duration_overflow_from_ndarray_containing_nat():
+    # GH52843
+    data_ts = pd.to_datetime([1, None])
+    data_td = pd.to_timedelta([1, None])
+    ser_ts = pd.Series(data_ts, dtype=ArrowDtype(pa.timestamp("ns")))
+    ser_td = pd.Series(data_td, dtype=ArrowDtype(pa.duration("ns")))
+    result = ser_ts + ser_td
+    expected = pd.Series([2, None], dtype=ArrowDtype(pa.timestamp("ns")))
+    tm.assert_series_equal(result, expected)