fix: Add fallback for incompatible string concenation in pandas

narwhals/_pandas_like/series.py

@@ -13,6 +13,7 @@
 from narwhals._pandas_like.utils import (
     NUMPY_VERSION,
     align_and_extract_native,
+    binary_string_sum_fallback,
     broadcast_series_to_index,
     get_dtype_backend,
     import_array_module,
@@ -399,23 +400,22 @@ def first(self) -> PythonLiteral:
     def last(self) -> PythonLiteral:
         return self.native.iloc[-1] if len(self.native) else None
 
-    def _with_binary(self, op: Callable[..., PandasLikeSeries], other: Any) -> Self:
+    def _with_binary(self, op: Callable[..., pd.Series], other: Any) -> Self:
         ser, other_native = align_and_extract_native(self, other)
         preserve_broadcast = self._broadcast and getattr(other, "_broadcast", True)
-        if (
-            str(self.native.dtype) == "large_string[pyarrow]"
-            and isinstance(other_native, str)
-            and op.__name__ == "add"
-        ):
-            # https://github.com/pandas-dev/pandas/issues/64393
-            import pyarrow as pa  # ignore-banned-import
-
-            other_native = pa.scalar(other_native, type=pa.large_string())
-        return self._with_native(
-            op(ser, other_native), preserve_broadcast=preserve_broadcast
-        ).alias(self.name)
+        try:
+            res = op(ser, other_native)
+        except Exception:
+            if op.__name__ == "add":
+                pdx = self.__native_namespace__()
+                res = binary_string_sum_fallback(ser, other_native, pdx)
+            else:
+                raise
+        return self._with_native(res, preserve_broadcast=preserve_broadcast).alias(
+            self.name
+        )
 
-    def _with_binary_right(self, op: Callable[..., PandasLikeSeries], other: Any) -> Self:
+    def _with_binary_right(self, op: Callable[..., pd.Series], other: Any) -> Self:
         return self._with_binary(lambda x, y: op(y, x), other).alias(self.name)
 
     def __eq__(self, other: object) -> Self:  # type: ignore[override]

narwhals/_pandas_like/utils.py

@@ -708,3 +708,36 @@ def broadcast_series_to_index(
         return series_class(pa_array, index=index, name=native.name)
 
     return series_class(value, index=index, dtype=native.dtype, name=native.name)
+
+
+def binary_string_sum_fallback(left: pd.Series, right: Any, pdx: Any) -> pd.Series:
+    # Workaround some upstream issues:
+    # - https://github.com/pandas-dev/pandas/issues/64393
+    # - https://github.com/pandas-dev/pandas/issues/65220
+    left_dtype = left.dtype
+    left_dtype_str = str(left_dtype)
+    if left_dtype_str == "large_string[pyarrow]" and isinstance(right, str):
+        import pyarrow as pa  # ignore-banned-import
+
+        return left + pa.scalar(right, type=pa.large_string())
+    if isinstance(right, pdx.Series):
+        right_dtype = right.dtype
+        if left_dtype_str == "object":  # pragma: no cover
+            # Only for pandas pre 3.0. Anything is better than `object`, so take RHS.
+            return left.astype(right_dtype) + right
+        if hasattr(left.values, "__arrow_array__") and hasattr(
+            right.values, "__arrow_array__"
+        ):
+            import pyarrow as pa  # ignore-banned-import
+
+            left_arrow = left.values.__arrow_array__().type  # noqa: PD011  # type: ignore[attr-defined]
+            right_arrow = right.values.__arrow_array__().type  # noqa: PD011  # type: ignore[attr-defined]
+            if pa.types.is_string(left_arrow) and pa.types.is_large_string(right_arrow):
+                # https://github.com/pandas-dev/pandas/blob/b00d4f6710ff6c1c80319196657c31c2cf6c70ff/pandas/core/arrays/arrow/array.py#L1064-L1068
+                pd_pa_large_string = pd.ArrowDtype(pa.large_string())
+                return left.astype(pd_pa_large_string) + right.astype(pd_pa_large_string)
+        else:  # pragma: no cover
+            pass
+        # Give precedence to the left-hand-side dtype.
+        return left + right.astype(left_dtype)
+    return left + right  # pragma: no cover

tests/expr_and_series/pandas_str_dtypes_test.py

@@ -0,0 +1,81 @@
+from __future__ import annotations
+
+from typing import Any
+
+import pytest
+
+import narwhals as nw
+from tests.utils import assert_equal_data
+
+pytest.importorskip("pandas", minversion="3.0.0")
+pytest.importorskip("pyarrow")
+
+import numpy as np
+import pandas as pd
+import pyarrow as pa
+
+STRING_DTYPE_NAN = pd.StringDtype("pyarrow", na_value=np.nan)  # type: ignore[call-arg]
+STRING_DTYPE_NA = pd.StringDtype("pyarrow", na_value=pd.NA)  # type: ignore[call-arg]
+
+
+@pytest.mark.parametrize(
+    ("left_dtype", "right_dtype", "result_dtype"),
+    [
+        (STRING_DTYPE_NAN, STRING_DTYPE_NAN, STRING_DTYPE_NAN),
+        (STRING_DTYPE_NAN, STRING_DTYPE_NA, STRING_DTYPE_NAN),
+        (STRING_DTYPE_NAN, pd.ArrowDtype(pa.string()), STRING_DTYPE_NAN),
+        (STRING_DTYPE_NAN, pd.ArrowDtype(pa.large_string()), STRING_DTYPE_NAN),
+        (STRING_DTYPE_NA, STRING_DTYPE_NAN, STRING_DTYPE_NA),
+        (STRING_DTYPE_NA, STRING_DTYPE_NA, STRING_DTYPE_NA),
+        (STRING_DTYPE_NA, pd.ArrowDtype(pa.string()), STRING_DTYPE_NA),
+        (STRING_DTYPE_NA, pd.ArrowDtype(pa.large_string()), STRING_DTYPE_NA),
+        (pd.ArrowDtype(pa.string()), STRING_DTYPE_NAN, pd.ArrowDtype(pa.large_string())),
+        (pd.ArrowDtype(pa.string()), STRING_DTYPE_NA, pd.ArrowDtype(pa.large_string())),
+        (
+            pd.ArrowDtype(pa.string()),
+            pd.ArrowDtype(pa.string()),
+            pd.ArrowDtype(pa.string()),
+        ),
+        (
+            pd.ArrowDtype(pa.string()),
+            pd.ArrowDtype(pa.large_string()),
+            pd.ArrowDtype(pa.large_string()),
+        ),
+        (
+            pd.ArrowDtype(pa.large_string()),
+            STRING_DTYPE_NAN,
+            pd.ArrowDtype(pa.large_string()),
+        ),
+        (
+            pd.ArrowDtype(pa.large_string()),
+            STRING_DTYPE_NA,
+            pd.ArrowDtype(pa.large_string()),
+        ),
+        (
+            pd.ArrowDtype(pa.large_string()),
+            pd.ArrowDtype(pa.string()),
+            pd.ArrowDtype(pa.large_string()),
+        ),
+        (
+            pd.ArrowDtype(pa.large_string()),
+            pd.ArrowDtype(pa.large_string()),
+            pd.ArrowDtype(pa.large_string()),
+        ),
+    ],
+)
+def test_pandas_str_types(left_dtype: Any, right_dtype: Any, result_dtype: Any) -> None:
+    import pandas as pd
+
+    df = pd.DataFrame({"fruit": ["apple", "banana"]}, dtype=left_dtype)
+    df["new_str_col"] = "!"
+    df["new_str_col"] = df["new_str_col"].astype(right_dtype)  # pyrefly: ignore[missing-attribute] https://github.com/facebook/pyrefly/issues/3299
+    res = nw.from_native(df).with_columns(
+        concat_col=nw.concat_str([nw.col("fruit"), nw.col("new_str_col")])
+    )
+    expected = {
+        "fruit": ["apple", "banana"],
+        "new_str_col": ["!", "!"],
+        "concat_col": ["apple!", "banana!"],
+    }
+    assert_equal_data(res, expected)
+    assert res.to_native()["concat_col"].dtype == result_dtype