BUG: lib.infer_dtype with incompatible intervals (pandas-dev#41749)

Author: jbrockmendel

pandas/_libs/lib.pyx

@@ -2031,16 +2031,59 @@ cdef bint is_period_array(ndarray[object] values):
     return True
 
 
-cdef class IntervalValidator(Validator):
-    cdef inline bint is_value_typed(self, object value) except -1:
-        return is_interval(value)
-
-
 cpdef bint is_interval_array(ndarray values):
+    """
+    Is this an ndarray of Interval (or np.nan) with a single dtype?
+    """
+
     cdef:
-        IntervalValidator validator = IntervalValidator(len(values),
-                                                        skipna=True)
-    return validator.validate(values)
+        Py_ssize_t i, n = len(values)
+        str closed = None
+        bint numeric = False
+        bint dt64 = False
+        bint td64 = False
+        object val
+
+    if len(values) == 0:
+        return False
+
+    for val in values:
+        if is_interval(val):
+            if closed is None:
+                closed = val.closed
+                numeric = (
+                    util.is_float_object(val.left)
+                    or util.is_integer_object(val.left)
+                )
+                td64 = is_timedelta(val.left)
+                dt64 = PyDateTime_Check(val.left)
+            elif val.closed != closed:
+                # mismatched closedness
+                return False
+            elif numeric:
+                if not (
+                        util.is_float_object(val.left)
+                        or util.is_integer_object(val.left)
+                    ):
+                    # i.e. datetime64 or timedelta64
+                    return False
+            elif td64:
+                if not is_timedelta(val.left):
+                    return False
+            elif dt64:
+                if not PyDateTime_Check(val.left):
+                    return False
+            else:
+                raise ValueError(val)
+        elif util.is_nan(val) or val is None:
+            pass
+        else:
+            return False
+
+    if closed is None:
+        # we saw all-NAs, no actual Intervals
+        return False
+    return True
 
 
 @cython.boundscheck(False)

pandas/core/construction.py

@@ -318,12 +318,7 @@ def array(
             return PeriodArray._from_sequence(data, copy=copy)
 
         elif inferred_dtype == "interval":
-            try:
-                return IntervalArray(data, copy=copy)
-            except ValueError:
-                # We may have a mixture of `closed` here.
-                # We choose to return an ndarray, rather than raising.
-                pass
+            return IntervalArray(data, copy=copy)
 
         elif inferred_dtype.startswith("datetime"):
             # datetime, datetime64

pandas/core/indexes/base.py

@@ -6443,12 +6443,8 @@ def _maybe_cast_data_without_dtype(subarr: np.ndarray) -> ArrayLike:
         return data
 
     elif inferred == "interval":
-        try:
-            ia_data = IntervalArray._from_sequence(subarr, copy=False)
-            return ia_data
-        except (ValueError, TypeError):
-            # GH27172: mixed closed Intervals --> object dtype
-            pass
+        ia_data = IntervalArray._from_sequence(subarr, copy=False)
+        return ia_data
     elif inferred == "boolean":
         # don't support boolean explicitly ATM
         pass

pandas/tests/dtypes/test_inference.py

@@ -1458,17 +1458,54 @@ def test_categorical(self):
         result = lib.infer_dtype(Series(arr), skipna=True)
         assert result == "categorical"
 
-    def test_interval(self):
+    @pytest.mark.parametrize("asobject", [True, False])
+    def test_interval(self, asobject):
         idx = pd.IntervalIndex.from_breaks(range(5), closed="both")
+        if asobject:
+            idx = idx.astype(object)
+
         inferred = lib.infer_dtype(idx, skipna=False)
         assert inferred == "interval"
 
         inferred = lib.infer_dtype(idx._data, skipna=False)
         assert inferred == "interval"
 
-        inferred = lib.infer_dtype(Series(idx), skipna=False)
+        inferred = lib.infer_dtype(Series(idx, dtype=idx.dtype), skipna=False)
         assert inferred == "interval"
 
+    @pytest.mark.parametrize("value", [Timestamp(0), Timedelta(0), 0, 0.0])
+    def test_interval_mismatched_closed(self, value):
+
+        first = Interval(value, value, closed="left")
+        second = Interval(value, value, closed="right")
+
+        # if closed match, we should infer "interval"
+        arr = np.array([first, first], dtype=object)
+        assert lib.infer_dtype(arr, skipna=False) == "interval"
+
+        # if closed dont match, we should _not_ get "interval"
+        arr2 = np.array([first, second], dtype=object)
+        assert lib.infer_dtype(arr2, skipna=False) == "mixed"
+
+    def test_interval_mismatched_subtype(self):
+        first = Interval(0, 1, closed="left")
+        second = Interval(Timestamp(0), Timestamp(1), closed="left")
+        third = Interval(Timedelta(0), Timedelta(1), closed="left")
+
+        arr = np.array([first, second])
+        assert lib.infer_dtype(arr, skipna=False) == "mixed"
+
+        arr = np.array([second, third])
+        assert lib.infer_dtype(arr, skipna=False) == "mixed"
+
+        arr = np.array([first, third])
+        assert lib.infer_dtype(arr, skipna=False) == "mixed"
+
+        # float vs int subdtype are compatible
+        flt_interval = Interval(1.5, 2.5, closed="left")
+        arr = np.array([first, flt_interval], dtype=object)
+        assert lib.infer_dtype(arr, skipna=False) == "interval"
+
     @pytest.mark.parametrize("klass", [pd.array, Series])
     @pytest.mark.parametrize("skipna", [True, False])
     @pytest.mark.parametrize("data", [["a", "b", "c"], ["a", "b", pd.NA]])