fix: improve validate columns (#829)

* improve validate columns

* skipif

* skipif

narwhals/_pandas_like/dataframe.py

@@ -1,6 +1,5 @@
 from __future__ import annotations
 
-import collections
 from typing import TYPE_CHECKING
 from typing import Any
 from typing import Iterable
@@ -26,6 +25,7 @@
 from narwhals.utils import parse_columns_to_drop
 
 if TYPE_CHECKING:
+    import pandas as pd
     from typing_extensions import Self
 
     from narwhals._pandas_like.group_by import PandasLikeGroupBy
@@ -73,15 +73,16 @@ def __native_namespace__(self) -> Any:
     def __len__(self) -> int:
         return len(self._native_frame)
 
-    def _validate_columns(self, columns: Sequence[str]) -> None:
-        if len(columns) != len(set(columns)):
-            counter = collections.Counter(columns)
-            for col, count in counter.items():
-                if count > 1:
-                    msg = f"Expected unique column names, got {col!r} {count} time(s)"
-                    raise ValueError(msg)
-            msg = "Please report a bug"  # pragma: no cover
-            raise AssertionError(msg)
+    def _validate_columns(self, columns: pd.Index) -> None:
+        try:
+            len_unique_columns = len(columns.drop_duplicates())
+        except Exception:  # noqa: BLE001  # pragma: no cover
+            msg = f"Expected hashable (e.g. str or int) column names, got: {columns}"
+            raise ValueError(msg) from None
+
+        if len(columns) != len_unique_columns:
+            msg = f"Expected unique column names, got: {columns}"
+            raise ValueError(msg)
 
     def _from_native_frame(self, df: Any) -> Self:
         return self.__class__(

tests/frame/schema_test.py

@@ -170,3 +170,31 @@ def test_unknown_dtype_polars() -> None:
 
 def test_hash() -> None:
     assert nw.Int64() in {nw.Int64, nw.Int32}
+
+
+@pytest.mark.parametrize(
+    ("method", "expected"),
+    [
+        ("names", ["a", "b", "c"]),
+        ("dtypes", [nw.Int64(), nw.Float32(), nw.String()]),
+        ("len", 3),
+    ],
+)
+def test_schema_object(method: str, expected: Any) -> None:
+    data = {"a": nw.Int64(), "b": nw.Float32(), "c": nw.String()}
+    schema = nw.Schema(data)
+    assert getattr(schema, method)() == expected
+
+
+@pytest.mark.skipif(
+    parse_version(pd.__version__) < (2,),
+    reason="Before 2.0, pandas would raise on `drop_duplicates`",
+)
+def test_from_non_hashable_column_name() -> None:
+    # This is technically super-illegal
+    # BUT, it shows up in a scikit-learn test, so...
+    df = pd.DataFrame([[1, 2], [3, 4]], columns=["pizza", ["a", "b"]])
+
+    df = nw.from_native(df, eager_only=True)
+    assert df.columns == ["pizza", ["a", "b"]]
+    assert df["pizza"].dtype == nw.Int64