diff --git a/src/databricks/labs/dqx/profiler/profiler.py b/src/databricks/labs/dqx/profiler/profiler.py
index 447b5549b..1cd604610 100644
--- a/src/databricks/labs/dqx/profiler/profiler.py
+++ b/src/databricks/labs/dqx/profiler/profiler.py
@@ -681,19 +681,29 @@ def _round_datetime(value: datetime.datetime, direction: str) -> datetime.dateti
         """
         Rounds a datetime value to midnight based on the specified direction.
 
+        - "down" → truncate to midnight (00:00:00).
+        - "up" → return the next midnight unless value is already midnight.
+        - Raises ValueError for invalid direction.
+
         :param value: The datetime value to round.
         :param direction: The rounding direction ("up" or "down").
         :return: The rounded datetime value.
+        :raises ValueError: If direction is not 'up' or 'down'.
         """
+        midnight = value.replace(hour=0, minute=0, second=0, microsecond=0)
+
         if direction == "down":
-            return value.replace(hour=0, minute=0, second=0, microsecond=0)
+            return midnight
+
         if direction == "up":
+            if midnight == value:
+                return value
             try:
-                return value.replace(hour=0, minute=0, second=0, microsecond=0) + datetime.timedelta(days=1)
+                return midnight + datetime.timedelta(days=1)
             except OverflowError:
                 logger.warning("Rounding datetime up caused overflow; returning datetime.max instead.")
                 return datetime.datetime.max
-        return value
+        raise ValueError(f"Invalid rounding direction: {direction}. Use 'up' or 'down'.")
 
     @staticmethod
     def _round_float(value: float, direction: str) -> float:
diff --git a/tests/integration/test_profiler.py b/tests/integration/test_profiler.py
index 733ef05c8..c842dccc4 100644
--- a/tests/integration/test_profiler.py
+++ b/tests/integration/test_profiler.py
@@ -104,6 +104,103 @@ def test_profiler(spark, ws):
     assert rules == expected_rules
 
 
+def test_profiler_rounding_midnight_behavior(spark, ws):
+    inp_schema = T.StructType(
+        [
+            T.StructField("t1", T.IntegerType()),
+            T.StructField("d1", T.DecimalType(10, 2)),
+            T.StructField("t2", T.StringType()),
+            T.StructField(
+                "s1",
+                T.StructType(
+                    [
+                        T.StructField("ns1", T.TimestampType()),
+                        T.StructField(
+                            "s2",
+                            T.StructType([T.StructField("ns2", T.StringType()), T.StructField("ns3", T.DateType())]),
+                        ),
+                    ]
+                ),
+            ),
+            T.StructField("b1", T.ByteType()),
+        ]
+    )
+    inp_df = spark.createDataFrame(
+        [
+            [
+                1,
+                Decimal("1.23"),
+                " test ",
+                {
+                    "ns1": datetime.fromisoformat("2023-01-08T00:00:00+00:00"),
+                    "s2": {"ns2": "test", "ns3": date.fromisoformat("2023-01-08")},
+                },
+                0,
+            ],
+            [
+                2,
+                Decimal("2.41"),
+                "test2",
+                {
+                    "ns1": datetime.fromisoformat("2023-01-07T10:00:11+00:00"),
+                    "s2": {"ns2": "test2", "ns3": date.fromisoformat("2023-01-07")},
+                },
+                1,
+            ],
+            [
+                3,
+                Decimal("333323.0"),
+                None,
+                {
+                    "ns1": datetime.fromisoformat("2023-01-06T10:00:11+00:00"),
+                    "s2": {"ns2": "test", "ns3": date.fromisoformat("2023-01-06")},
+                },
+                0,
+            ],
+        ],
+        schema=inp_schema,
+    )
+
+    profiler = DQProfiler(ws)
+    stats, rules = profiler.profile(inp_df, options={"sample_fraction": None})
+
+    expected_rules = [
+        DQProfile(name="is_not_null", column="t1", description=None, parameters=None),
+        DQProfile(
+            name="min_max", column="t1", description="Real min/max values were used", parameters={"min": 1, "max": 3}
+        ),
+        DQProfile(name='is_not_null', column='d1', description=None, parameters=None),
+        DQProfile(
+            name='min_max',
+            column='d1',
+            description='Real min/max values were used',
+            parameters={'max': Decimal('333323.00'), 'min': Decimal('1.23')},
+        ),
+        DQProfile(name='is_not_null_or_empty', column='t2', description=None, parameters={'trim_strings': True}),
+        DQProfile(name="is_not_null", column="s1.ns1", description=None, parameters=None),
+        DQProfile(
+            name="min_max",
+            column="s1.ns1",
+            description="Real min/max values were used",
+            parameters={
+                "min": datetime(2023, 1, 6, 0, 0, tzinfo=timezone.utc),
+                "max": datetime(2023, 1, 8, 0, 0, tzinfo=timezone.utc),
+            },
+        ),
+        DQProfile(name="is_not_null", column="s1.s2.ns2", description=None, parameters=None),
+        DQProfile(name="is_not_null", column="s1.s2.ns3", description=None, parameters=None),
+        DQProfile(
+            name="min_max",
+            column="s1.s2.ns3",
+            description="Real min/max values were used",
+            parameters={"min": date(2023, 1, 6), "max": date(2023, 1, 8)},
+        ),
+        DQProfile(name="is_not_null", column="b1", description=None, parameters=None),
+    ]
+    assert len(stats.keys()) > 0
+    assert rules == expected_rules
+
+
 def test_profiler_non_default_profile_options(spark, ws):
     inp_schema = T.StructType(
         [