Bug: to_csv coercing datetime columns to same format

doc/source/whatsnew/v1.4.0.rst

@@ -758,6 +758,7 @@ I/O
 - Bug in :func:`read_csv` converting columns to numeric after date parsing failed (:issue:`11019`)
 - Bug in :func:`read_csv` not replacing ``NaN`` values with ``np.nan`` before attempting date conversion (:issue:`26203`)
 - Bug in :func:`read_csv` raising ``AttributeError`` when attempting to read a .csv file and infer index column dtype from an nullable integer type (:issue:`44079`)
+- Bug in :func:`to_csv` always coercing datetime columns with different formats to the same format (:issue:`21734`)
 - :meth:`DataFrame.to_csv` and :meth:`Series.to_csv` with ``compression`` set to ``'zip'`` no longer create a zip file containing a file ending with ".zip". Instead, they try to infer the inner file name more smartly. (:issue:`39465`)
 - Bug in :func:`read_csv` where reading a mixed column of booleans and missing values to a float type results in the missing values becoming 1.0 rather than NaN (:issue:`42808`, :issue:`34120`)
 - Bug in :func:`read_csv` when passing simultaneously a parser in ``date_parser`` and ``parse_dates=False``, the parsing was still called (:issue:`44366`)

pandas/core/internals/blocks.py

@@ -2103,9 +2103,12 @@ def to_native_types(
     values = ensure_wrapped_if_datetimelike(values)
 
     if isinstance(values, (DatetimeArray, TimedeltaArray)):
-        result = values._format_native_types(na_rep=na_rep, **kwargs)
-        result = result.astype(object, copy=False)
-        return result
+        # GH#21734 Process every column separate, they migh have different formats
+        results_converted = []
+        for i in range(len(values)):
+            result = values[i, :]._format_native_types(na_rep=na_rep, **kwargs)
+            results_converted.append(result.astype(object, copy=False))
+        return np.vstack(results_converted)
 
     elif isinstance(values, ExtensionArray):
         mask = isna(values)

pandas/tests/io/formats/test_to_csv.py

@@ -282,6 +282,22 @@ def test_to_csv_date_format(self):
         df_sec_grouped = df_sec.groupby([pd.Grouper(key="A", freq="1h"), "B"])
         assert df_sec_grouped.mean().to_csv(date_format="%Y-%m-%d") == expected_ymd_sec
 
+    def test_to_csv_different_datetime_formats(self):
+        # GH#21734
+        df = DataFrame(
+            {
+                "date": pd.to_datetime("1970-01-01"),
+                "datetime": pd.date_range("1970-01-01", periods=2, freq="H"),
+            }
+        )
+        expected_rows = [
+            "date,datetime",
+            "1970-01-01,1970-01-01 00:00:00",
+            "1970-01-01,1970-01-01 01:00:00",
+        ]
+        expected = tm.convert_rows_list_to_csv_str(expected_rows)
+        assert df.to_csv(index=False) == expected
+
     def test_to_csv_multi_index(self):
         # see gh-6618
         df = DataFrame([1], columns=pd.MultiIndex.from_arrays([[1], [2]]))