ntkathole · Apr 14, 2026 · Apr 10, 2026 · Apr 14, 2026
@@ -909,6 +909,16 @@ def _convert_list_values_to_proto(
     ]
 
 
+def _is_array_like(value: Any) -> bool:
+    """Return True if *value* is array-like (numpy array or any sized,
+    non-string, non-bytes container).  Array-like values in a scalar
+    feature column cannot be mapped to a protobuf scalar field and are
+    therefore always treated as null."""
+    return isinstance(value, np.ndarray) or (
+        hasattr(value, "__len__") and not isinstance(value, (str, bytes))
+    )
+
+
 def _convert_scalar_values_to_proto(
     feast_value_type: ValueType,
     values: List[Any],
@@ -929,16 +939,34 @@ def _convert_scalar_values_to_proto(
         return [ProtoValue()] * len(values)
 
     if feast_value_type == ValueType.UNIX_TIMESTAMP:
-        int_timestamps = _python_datetime_to_int_timestamp(values)
-        return [ProtoValue(unix_timestamp_val=ts) for ts in int_timestamps]  # type: ignore
+        out: List[Any] = [None] * len(values)
+        clean_indices: List[int] = []
+        clean_values: List[Any] = []
+        for i, value in enumerate(values):
+            if _is_array_like(value) or value is None:
+                out[i] = ProtoValue()
+            else:
+                clean_indices.append(i)
+                clean_values.append(value)
+        if clean_values:
+            timestamps = _python_datetime_to_int_timestamp(clean_values)
+            for i, ts in zip(clean_indices, timestamps):
+                out[i] = ProtoValue(unix_timestamp_val=ts)  # type: ignore
+        return out
 
     field_name, func, valid_scalar_types = PYTHON_SCALAR_VALUE_TYPE_TO_PROTO_VALUE[
         feast_value_type
     ]
 
-    # Validate scalar types
+    # Validate scalar types.  The caller guarantees that *sample* is not
+    # array-like (array-like values are filtered out when picking the sample
+    # for scalar columns in python_values_to_proto_values).
     if valid_scalar_types:
-        if (sample == 0 or sample == 0.0) and feast_value_type != ValueType.BOOL:
+        try:
+            is_zero = sample == 0 or sample == 0.0
+        except (ValueError, TypeError):
+            is_zero = False
+        if is_zero and feast_value_type != ValueType.BOOL:
             # Numpy converts 0 to int, but column type may be float
             allowed_types = {np.int64, int, np.float64, float, decimal.Decimal}
             assert type(sample) in allowed_types, (
@@ -951,20 +979,35 @@ def _convert_scalar_values_to_proto(
 
     # Handle BOOL specially due to np.bool_ conversion requirement
     if feast_value_type == ValueType.BOOL:
-        return [
-            ProtoValue(
-                **{field_name: func(bool(value) if type(value) is np.bool_ else value)}
-            )  # type: ignore
-            if not pd.isnull(value)
-            else ProtoValue()
-            for value in values
-        ]
+        out = []
+        for value in values:
+            if _is_array_like(value):
+                # Array-like value in a scalar BOOL column: treat as null.
+                out.append(ProtoValue())
+            elif not pd.isnull(value):
+                out.append(
+                    ProtoValue(
+                        **{
+                            field_name: func(
+                                bool(value) if type(value) is np.bool_ else value
+                            )
+                        }
+                    )  # type: ignore
+                )
+            else:
+                out.append(ProtoValue())
+        return out
 
     # Generic scalar conversion
     out = []
     for value in values:
         if isinstance(value, ProtoValue):
             out.append(value)
+        elif _is_array_like(value):
+            # Array-like value in a scalar column: always treat as null.
+            # pd.isnull() is vectorised and would return an ndarray here,
+            # making `not pd.isnull(value)` raise ValueError.
+            out.append(ProtoValue())
         elif not pd.isnull(value):
             out.append(ProtoValue(**{field_name: func(value)}))
         else:
@@ -1107,12 +1150,18 @@ def _python_value_to_proto_value(
     if "set" in type_name_lower:
         return _python_set_to_proto_values(feast_value_type, values)
 
-    # Scalar types
+    # Scalar types — pick a sample that is not array-like so that the type
+    # validation in _convert_scalar_values_to_proto always receives a plain
+    # scalar (array-like values in a scalar column are treated as null).
     if (
         feast_value_type in PYTHON_SCALAR_VALUE_TYPE_TO_PROTO_VALUE
         or feast_value_type == ValueType.UNIX_TIMESTAMP
     ):
-        return _convert_scalar_values_to_proto(feast_value_type, values, sample)
+        scalar_sample = next(
+            (v for v in values if _non_empty_value(v) and not _is_array_like(v)),
+            None,
+        )
+        return _convert_scalar_values_to_proto(feast_value_type, values, scalar_sample)
 
     raise Exception(f"Unsupported data type: {feast_value_type}")
 

@@ -1845,3 +1845,111 @@ def test_pa_to_feast_value_type_nested(self):
             pa_to_feast_value_type("list<item: list<item: double>>")
             == ValueType.VALUE_LIST
         )
+
+
+class TestEmptyArrayAsNull:
+    """Regression tests for https://github.com/feast-dev/feast/issues/6255
+    Ensure that an empty numpy array in a scalar feature column is treated as
+    null rather than raising ``ValueError: The truth value of an empty array is
+    ambiguous``.
+    """
+
+    def test_empty_numpy_array_treated_as_null_double(self):
+        from feast.protos.feast.types.Value_pb2 import Value as ProtoValue
+
+        result = python_values_to_proto_values(
+            [np.array([]), 1.0, None], ValueType.DOUBLE
+        )
+        assert result[0] == ProtoValue(), (
+            "empty array should produce an empty ProtoValue"
+        )
+        assert result[1].double_val == 1.0
+        assert result[2] == ProtoValue(), (
+            "None should still produce an empty ProtoValue"
+        )
+
+    def test_empty_numpy_array_treated_as_null_int64(self):
+        from feast.protos.feast.types.Value_pb2 import Value as ProtoValue
+
+        result = python_values_to_proto_values(
+            [np.array([]), 42, None], ValueType.INT64
+        )
+        assert result[0] == ProtoValue(), (
+            "empty array should produce an empty ProtoValue"
+        )
+        assert result[1].int64_val == 42
+        assert result[2] == ProtoValue()
+
+    def test_empty_numpy_array_treated_as_null_bool(self):
+        from feast.protos.feast.types.Value_pb2 import Value as ProtoValue
+
+        result = python_values_to_proto_values(
+            [np.array([]), True, None], ValueType.BOOL
+        )
+        assert result[0] == ProtoValue(), (
+            "empty array should produce an empty ProtoValue"
+        )
+        assert result[1].bool_val is True
+        assert result[2] == ProtoValue()
+
+    def test_array_with_null_element_treated_as_null(self):
+        """A non-empty array containing any null element in a scalar column is treated as null."""
+        from feast.protos.feast.types.Value_pb2 import Value as ProtoValue
+
+        result = python_values_to_proto_values(
+            [np.array([np.nan, 1.0]), 3.0], ValueType.DOUBLE
+        )
+        assert result[0] == ProtoValue(), (
+            "array with null element should produce an empty ProtoValue"
+        )
+        assert result[1].double_val == 3.0
+
+    def test_non_empty_array_without_nulls_is_treated_as_null(self):
+        """A non-empty numpy array in a scalar column is always treated as null.
+
+        A scalar feature column cannot hold an ndarray value (protobuf would
+        reject it), so any array-like value – empty or not – is mapped to an
+        empty ProtoValue() rather than crashing with ValueError.
+        """
+        from feast.protos.feast.types.Value_pb2 import Value as ProtoValue
+
+        result = python_values_to_proto_values(
+            [np.array([1.0, 2.0]), 3.0, None], ValueType.DOUBLE
+        )
+        # array-like value in a scalar column → null, not a crash
+        assert result[0] == ProtoValue(), (
+            "non-empty array in scalar column should be null"
+        )
+        assert result[1].double_val == 3.0
+        assert result[2] == ProtoValue()
+
+    def test_empty_numpy_array_treated_as_null_unix_timestamp(self):
+        """Array-like values in a scalar UNIX_TIMESTAMP column must not crash."""
+        from datetime import datetime, timezone
+
+        from feast.protos.feast.types.Value_pb2 import Value as ProtoValue
+
+        ts = datetime(2024, 1, 1, tzinfo=timezone.utc)
+        result = python_values_to_proto_values(
+            [np.array([]), ts, None], ValueType.UNIX_TIMESTAMP
+        )
+        assert result[0] == ProtoValue(), (
+            "empty array in UNIX_TIMESTAMP scalar column should produce null"
+        )
+        assert result[1].unix_timestamp_val == int(ts.timestamp())
+        assert result[2] == ProtoValue()
+
+    def test_non_empty_array_treated_as_null_unix_timestamp(self):
+        """Non-empty array in a UNIX_TIMESTAMP scalar column should produce null, not crash."""
+        from datetime import datetime, timezone
+
+        from feast.protos.feast.types.Value_pb2 import Value as ProtoValue
+
+        ts = datetime(2024, 6, 15, tzinfo=timezone.utc)
+        result = python_values_to_proto_values(
+            [np.array([1, 2, 3]), ts], ValueType.UNIX_TIMESTAMP
+        )
+        assert result[0] == ProtoValue(), (
+            "non-empty array in UNIX_TIMESTAMP scalar column should produce null"
+        )
+        assert result[1].unix_timestamp_val == int(ts.timestamp())