fix: Allow overriding column definitions in schema inheritance

dataframely/_base_schema.py

@@ -119,7 +119,9 @@ def __new__(
         result = Metadata()
         for base in bases:
             result.update(mcs._get_metadata_recursively(base))
-        result.update(mcs._get_metadata(namespace))
+        namespace_metadata = mcs._get_metadata(namespace)
+        mcs._remove_overridden_columns(result, namespace, bases)
+        result.update(namespace_metadata)
         namespace[_COLUMN_ATTR] = result.columns
         cls = super().__new__(mcs, name, bases, namespace, *args, **kwargs)
 
@@ -207,6 +209,34 @@ def __getattribute__(cls, name: str) -> Any:
                 val._name = val.alias or name
             return val
 
+    @staticmethod
+    def _remove_overridden_columns(
+        result: Metadata,
+        namespace: dict[str, Any],
+        bases: tuple[type[object], ...],
+    ) -> None:
+        """Remove inherited columns that the child namespace explicitly overrides.
+
+        Before merging the child namespace, we must drop any parent columns whose
+        attribute name is redefined in the child. This allows subclasses to redefine
+        inherited columns while still detecting genuine alias conflicts.
+
+        In multiple-inheritance scenarios, the same attribute name may appear in more
+        than one base with different aliases, so we walk all parent MROs and collect
+        every matching column key to remove.
+        """
+        for attr, value in namespace.items():
+            if not isinstance(value, Column):
+                continue
+            keys_to_remove: set[str] = set()
+            for base in bases:
+                for parent_cls in base.__mro__:
+                    parent_col = parent_cls.__dict__.get(attr)
+                    if parent_col is not None and isinstance(parent_col, Column):
+                        keys_to_remove.add(parent_col.alias or attr)
+            for parent_key in keys_to_remove:
+                result.columns.pop(parent_key, None)
+
     @staticmethod
     def _get_metadata_recursively(kls: type[object]) -> Metadata:
         result = Metadata()

tests/schema/test_base.py

@@ -141,3 +141,22 @@ def test_user_error_polars_datatype_type() -> None:
         class MySchemaWithPolarsDataTypeType(dy.Schema):
             a = dy.Int32(nullable=False)
             b = pl.String  # User error: Used pl.String instead of dy.String()
+
+
+def test_override() -> None:
+    class FirstSchema(dy.Schema):
+        x = dy.Int64()
+
+    class SecondSchema(FirstSchema):
+        x = dy.Int64(nullable=True)
+
+    first_columns = FirstSchema.columns()
+    second_columns = SecondSchema.columns()
+
+    assert set(first_columns) == {"x"}
+    assert set(second_columns) == {"x"}
+
+    assert first_columns["x"].nullable is False
+    assert second_columns["x"].nullable is True
+
+    assert type(second_columns["x"]) is type(first_columns["x"])