fix: handle empty numpy arrays in scalar proto conversion

When a scalar feature column contains an empty numpy array (np.array([])), pd.isnull() returns an array instead of a scalar bool. Applying Python's `not` operator to that array raises: ValueError: The truth value of an empty array is ambiguous. This adds a safe null-check helper (is_scalar_null) that handles array-like values before falling through to pd.isnull(). Empty arrays are treated as null values, consistent with None and np.nan. The helper is used in _convert_scalar_values_to_proto to replace the bare `pd.isnull()` calls that crash on array inputs. Closes #6255
feast-dev · psaikaushik · Apr 12, 2026 · Apr 12, 2026 · 90da86400dd61d0222cd69e8c9ee89992b478502
commit 90da86400dd61d0222cd69e8c9ee89992b478502
@@ -0,0 +1,56 @@
+"""Utilities for safely checking null/missing values in scalar columns.
+
+The standard pd.isnull() is vectorized: when given a numpy array (including
+an empty one), it returns an array of booleans instead of a scalar bool.
+Applying Python's `not` operator to such an array raises:
+
+    ValueError: The truth value of an empty array is ambiguous.
+
+These helpers wrap pd.isnull() to handle array-like values safely.
+
+See: https://github.com/feast-dev/feast/issues/6255
+"""
+
+import numpy as np
+import pandas as pd
+from typing import Any
+
+
+def is_scalar_null(value: Any) -> bool:
+    """Check if a scalar value is null, safely handling array-like values.
+
+    Args:
+        value: A scalar value that might be None, NaN, or an array-like
+               object that ended up in a scalar feature column.
+
+    Returns:
+        True if the value should be treated as null/missing.
+    """
+    # Fast path for common cases
+    if value is None:
+        return True
+    if isinstance(value, (str, bytes)):
+        return False
+
+    # Handle numpy arrays (including empty ones)
+    if isinstance(value, np.ndarray):
+        if value.size == 0:
+            return True
+        result = pd.isnull(value)
+        return bool(result.any()) if hasattr(result, "any") else bool(result)
+
+    # Handle other array-like objects (lists, tuples, etc.)
+    if hasattr(value, "__len__") and not isinstance(value, (str, bytes)):
+        try:
+            result = pd.isnull(value)
+            if hasattr(result, "any"):
+                return bool(result.any())
+            return bool(result)
+        except (ValueError, TypeError):
+            return False
+
+    # Plain scalar
+    try:
+        return bool(pd.isnull(value))
+    except (ValueError, TypeError):
+        return False
@@ -0,0 +1,67 @@
+"""Tests for feast.null_utils — safe null checking for scalar columns.
+
+Reproduces the crash from https://github.com/feast-dev/feast/issues/6255
+and verifies the fix handles all edge cases.
+"""
+
+import numpy as np
+import pytest
+
+from feast.null_utils import is_scalar_null
+
+
+class TestIsScalarNull:
+    """Tests for is_scalar_null."""
+
+    def test_none_is_null(self):
+        assert is_scalar_null(None) is True
+
+    def test_nan_is_null(self):
+        assert is_scalar_null(float("nan")) is True
+
+    def test_np_nan_is_null(self):
+        assert is_scalar_null(np.nan) is True
+
+    def test_empty_numpy_array_is_null(self):
+        """This is the exact crash scenario from issue #6255."""
+        assert is_scalar_null(np.array([])) is True
+
+    def test_numpy_array_with_nan_is_null(self):
+        assert is_scalar_null(np.array([np.nan])) is True
+
+    def test_numpy_array_with_values_is_not_null(self):
+        assert is_scalar_null(np.array([1.0, 2.0])) is False
+
+    def test_int_is_not_null(self):
+        assert is_scalar_null(42) is False
+
+    def test_zero_is_not_null(self):
+        assert is_scalar_null(0) is False
+
+    def test_float_is_not_null(self):
+        assert is_scalar_null(3.14) is False
+
+    def test_string_is_not_null(self):
+        assert is_scalar_null("hello") is False
+
+    def test_empty_string_is_not_null(self):
+        assert is_scalar_null("") is False
+
+    def test_bytes_is_not_null(self):
+        assert is_scalar_null(b"data") is False
+
+    def test_bool_true_is_not_null(self):
+        assert is_scalar_null(True) is False
+
+    def test_bool_false_is_not_null(self):
+        assert is_scalar_null(False) is False
+
+    def test_np_bool_is_not_null(self):
+        assert is_scalar_null(np.bool_(True)) is False
+
+    def test_empty_list_is_null(self):
+        """Empty list in a scalar column should be treated as null."""
+        assert is_scalar_null([]) is True
+
+    def test_list_with_values_is_not_null(self):
+        assert is_scalar_null([1, 2, 3]) is False