Add test for concat datetime64 columns of differing resolutions

pandas/tests/reshape/concat/test_datetimes.py

@@ -278,6 +278,32 @@ def test_concat_NaT_series_dataframe_all_NaT(self, tz1, tz2):
         result = concat([first, second])
         tm.assert_frame_equal(result, expected)
 
+    def test_concat_ns_and_s_preserves_datetime64(self):
+        # GH 53307
+        # ensure concatenating a datetime64[ns] column and a copy cast to M8[s]
+        # yields a datetime64 dtype (finest unit should be ns)
+        df = pd.DataFrame(
+            {"ints": range(2), "dates": pd.date_range("2000", periods=2, freq="min")}
+        )
+        df2 = df.copy()
+        df2["dates"] = df2["dates"].astype("M8[s]")
+
+        combined = pd.concat([df, df2], ignore_index=True)
+
+        # dtype is a datetime64 type
+        assert pd.api.types.is_datetime64_any_dtype(combined["dates"].dtype)
+
+        # unit should be the finest (ns) when mixing ns and s
+        unit = np.datetime_data(combined["dates"].dtype)[0]
+        assert unit == "ns"
+
+        # values preserved (compare as ns)
+        exp = pd.to_datetime(list(df["dates"]) + list(df2["dates"]))
+        tm.assert_series_equal(
+            combined["dates"].astype("datetime64[ns]").reset_index(drop=True),
+            pd.Series(exp.astype("datetime64[ns]"), name="dates").reset_index(drop=True),
+        )
+
 
 class TestTimezoneConcat:
     def test_concat_tz_series(self):
@@ -591,3 +617,52 @@ def test_concat_float_datetime64():
 
     result = concat([df_time, df_float.iloc[:0]])
     tm.assert_frame_equal(result, expected)
+
+
+@pytest.mark.parametrize(
+    "unit,unit2",
+    [(u1, u2) for u1 in ("ns", "us", "ms", "s") for u2 in ("ns", "us", "ms", "s")],
+)
+def test_concat_mixed_units_preserve_datetime_and_unit(unit, unit2):
+    # GH 53307
+    # for each pair of units, concatenating columns of those units should
+    # result in a datetime64 dtype with the finest unit
+    df = pd.DataFrame({"dates": pd.to_datetime(["2000-01-01", "2000-01-02"])})
+    # cast copies to requested unit
+    df1 = df.copy()
+    df1["dates"] = df1["dates"].astype(f"M8[{unit}]")
+    df2 = df.copy()
+    df2["dates"] = df2["dates"].astype(f"M8[{unit2}]")
+
+    exp_unit = tm.get_finest_unit(unit, unit2)
+
+    # test both concat orders
+    for a, b in ((df1, df2), (df2, df1)):
+        combined = pd.concat([a, b], ignore_index=True)
+
+        assert pd.api.types.is_datetime64_any_dtype(combined["dates"].dtype)
+
+        res_unit = np.datetime_data(combined["dates"].dtype)[0]
+        assert res_unit == exp_unit
+
+
+@pytest.mark.parametrize(
+    "unit,unit2",
+    [(u1, u2) for u1 in ("ns", "us", "ms", "s") for u2 in ("ns", "us", "ms", "s")],
+)
+def test_concat_mixed_units_with_all_nat(unit, unit2):
+    # GH 53307
+    # mixing non-empty datetime column and an all-NaT column typed to unit2
+    df = pd.DataFrame({"dates": pd.to_datetime(["2000-01-01"])})
+    df1 = df.copy()
+    df1["dates"] = df1["dates"].astype(f"M8[{unit}]")
+
+    ser_nat = pd.Series([pd.NaT], dtype=f"datetime64[{unit2}]")
+    df2 = pd.DataFrame({"dates": ser_nat})
+
+    exp_unit = tm.get_finest_unit(unit, unit2)
+
+    combined = pd.concat([df1, df2], ignore_index=True)
+    assert pd.api.types.is_datetime64_any_dtype(combined["dates"].dtype)
+    res_unit = np.datetime_data(combined["dates"].dtype)[0]
+    assert res_unit == exp_unit