fixing bug

Author: manu-sj

python/hsfs/feature_view.py

@@ -5111,12 +5111,19 @@ def on_demand_transformations(self) -> dict[str, Callable]:
 
     @property
     def _on_demand_transformation_functions(self) -> list[TransformationFunction]:
-        """Get all on-demand transformations in the feature view."""
-        return [
-            feature.on_demand_transformation_function
-            for feature in self.features
-            if feature.on_demand_transformation_function
-        ]
+        """Get all on-demand transformations in the feature view.
+
+        A multi-output TF appears on every feature it produces, so the raw
+        collection contains duplicates. Dedup by `__eq__` (which compares the
+        serialized TF) and preserve the first occurrence so downstream DAG
+        construction sees each TF exactly once.
+        """
+        seen: list[TransformationFunction] = []
+        for feature in self.features:
+            tf = feature.on_demand_transformation_function
+            if tf and tf not in seen:
+                seen.append(tf)
+        return seen
 
     @public
     @property

python/tests/test_feature_view.py

@@ -162,6 +162,65 @@ def test_from_response_json_transformation_function(self, mocker, backend_fixtur
         assert len(fv.schema) == 2
         assert isinstance(fv.schema[0], training_dataset_feature.TrainingDatasetFeature)
 
+    def test_on_demand_transformation_functions_dedup_multi_output_tf(self, mocker):
+        """Regression: a multi-output ODT is attached to each feature it produces.
+
+        Without dedup, the property returns the TF twice — once per feature — and
+        TransformationExecutionDAG raises "Output column 'x' is produced by both
+        'tf' and 'tf'." This was hit at FV creation time when reading the
+        backend response back, breaking any chained pipeline with a multi-output
+        TF (e.g. add_two -> odt2_1, odt2_2).
+        """
+        from hsfs.core import transformation_execution_dag
+        from hsfs.transformation_function import (
+            TransformationFunction,
+            TransformationType,
+        )
+
+        mocker.patch("hopsworks_common.client.get_instance")
+        mocker.patch("hsfs.engine.get_type", return_value="python")
+
+        @udf([float, float])
+        def add_two(feature):
+            return feature + 2, feature + 2
+
+        tf = TransformationFunction(
+            featurestore_id=99,
+            hopsworks_udf=add_two,
+            transformation_type=TransformationType.ON_DEMAND,
+        )("feature").alias("odt2_1", "odt2_2")
+
+        fv = feature_view.FeatureView(
+            name="fv_dedup",
+            query=fg1.select_features(),
+            featurestore_id=99,
+            featurestore_name="test_fs",
+        )
+        # Multi-output TF -> the SAME TF appears on every output feature.
+        # The backend response path produces fresh Python instances per feature,
+        # so simulate that by attaching distinct copies that compare equal.
+        fv.schema = [
+            training_dataset_feature.TrainingDatasetFeature(
+                name="odt2_1",
+                type="float",
+                transformation_function=tf,
+            ),
+            training_dataset_feature.TrainingDatasetFeature(
+                name="odt2_2",
+                type="float",
+                transformation_function=tf,
+            ),
+        ]
+
+        # Property must dedup
+        odts = fv._on_demand_transformation_functions
+        assert len(odts) == 1
+        assert odts[0].hopsworks_udf.function_name == "add_two"
+
+        # DAG construction must succeed (no "produced by both 'add_two' and 'add_two'")
+        dag = transformation_execution_dag.TransformationExecutionDAG(odts)
+        assert len(dag.nodes) == 1
+
     def test_from_response_json_basic_info_deprecated(self, mocker, backend_fixtures):
         # Arrange
         mocker.patch("hsfs.engine.get_type")