BUG: tz_localize needs to invalidate freq

doc/source/whatsnew/v1.1.0.rst

@@ -412,7 +412,7 @@ Datetimelike
 - Bug where :meth:`PeriodIndex` raised when passed a :class:`Series` of strings (:issue:`26109`)
 - Bug in :class:`Timestamp` arithmetic when adding or subtracting a ``np.ndarray`` with ``timedelta64`` dtype (:issue:`33296`)
 - Bug in :meth:`DatetimeIndex.to_period` not infering the frequency when called with no arguments (:issue:`33358`)
-
+- Bug in :meth:`DatetimeIndex.tz_localize` incorrectly retaining ``freq`` in some cases where the original freq is no longer valid (:issue:`30511`)
 
 Timedelta
 ^^^^^^^^^

pandas/core/arrays/datetimes.py

@@ -886,15 +886,15 @@ def tz_localize(self, tz, ambiguous="raise", nonexistent="raise"):
         DatetimeIndex(['2018-03-01 09:00:00-05:00',
                        '2018-03-02 09:00:00-05:00',
                        '2018-03-03 09:00:00-05:00'],
-                      dtype='datetime64[ns, US/Eastern]', freq='D')
+                      dtype='datetime64[ns, US/Eastern]', freq=None)
 
         With the ``tz=None``, we can remove the time zone information
         while keeping the local time (not converted to UTC):
 
         >>> tz_aware.tz_localize(None)
         DatetimeIndex(['2018-03-01 09:00:00', '2018-03-02 09:00:00',
                        '2018-03-03 09:00:00'],
-                      dtype='datetime64[ns]', freq='D')
+                      dtype='datetime64[ns]', freq=None)
 
         Be careful with DST changes. When there is sequential data, pandas can
         infer the DST time:
@@ -973,7 +973,16 @@ def tz_localize(self, tz, ambiguous="raise", nonexistent="raise"):
             )
         new_dates = new_dates.view(DT64NS_DTYPE)
         dtype = tz_to_dtype(tz)
-        return self._simple_new(new_dates, dtype=dtype, freq=self.freq)
+
+        freq = None
+        if timezones.is_utc(tz) or (len(self) == 1 and not isna(new_dates[0])):
+            # we can preserve freq
+            # TODO: Also for fixed-offsets
+            freq = self.freq
+        elif tz is None and self.tz is None:
+            # no-op
+            freq = self.freq
+        return self._simple_new(new_dates, dtype=dtype, freq=freq)
 
     # ----------------------------------------------------------------
     # Conversion Methods - Vectorized analogues of Timestamp methods

pandas/tests/indexes/datetimes/test_timezones.py

@@ -1161,3 +1161,26 @@ def test_iteration_preserves_nanoseconds(self, tz):
         )
         for i, ts in enumerate(index):
             assert ts == index[i]
+
+
+def test_tz_localize_invalidates_freq():
+    # we only preserve freq in unambiguous cases
+
+    # if localized to US/Eastern, this crosses a DST transition
+    dti = date_range("2014-03-08 23:00", "2014-03-09 09:00", freq="H")
+    assert dti.freq == "H"
+
+    result = dti.tz_localize(None)  # no-op
+    assert result.freq == "H"
+
+    result = dti.tz_localize("UTC")  # unambiguous freq preservation
+    assert result.freq == "H"
+
+    result = dti.tz_localize("US/Eastern", nonexistent="shift_forward")
+    assert result.freq is None
+    assert result.inferred_freq is None  # i.e. we are not _too_ strict here
+
+    # Case where we _can_ keep freq because we're length==1
+    dti2 = dti[:1]
+    result = dti2.tz_localize("US/Eastern")
+    assert result.freq == "H"

pandas/tests/series/test_arithmetic.py

@@ -378,6 +378,7 @@ def test_ser_cmp_result_names(self, names, op):
 
         # datetime64tz dtype
         dti = dti.tz_localize("US/Central")
+        dti._set_freq("infer")  # freq not preserved by tz_localize
         ser = Series(dti).rename(names[1])
         result = op(ser, dti)
         assert result.name == names[2]