ENH: infer Timestamp(iso8601string) resolution (#49737)

* ENH: infer Timestamp(iso8601string) resolution

* xfail windwos

* fix assertion

* post-merge fixup

* Update doc/source/whatsnew/v2.0.0.rst

Co-authored-by: Matthew Roeschke <10647082+mroeschke@users.noreply.github.com>

* CLN: remove commented-out

* update test

* test for unit on parsing iso8601 with offset

* test for round

* fix offset tests

Co-authored-by: Matthew Roeschke <10647082+mroeschke@users.noreply.github.com>

Author: jbrockmendel

doc/source/whatsnew/v2.0.0.rst

@@ -465,7 +465,8 @@ Other API changes
 - :meth:`Index.astype` now allows casting from ``float64`` dtype to datetime-like dtypes, matching :class:`Series` behavior (:issue:`49660`)
 - Passing data with dtype of "timedelta64[s]", "timedelta64[ms]", or "timedelta64[us]" to :class:`TimedeltaIndex`, :class:`Series`, or :class:`DataFrame` constructors will now retain that dtype instead of casting to "timedelta64[ns]"; timedelta64 data with lower resolution will be cast to the lowest supported resolution "timedelta64[s]" (:issue:`49014`)
 - Passing ``dtype`` of "timedelta64[s]", "timedelta64[ms]", or "timedelta64[us]" to :class:`TimedeltaIndex`, :class:`Series`, or :class:`DataFrame` constructors will now retain that dtype instead of casting to "timedelta64[ns]"; passing a dtype with lower resolution for :class:`Series` or :class:`DataFrame` will be cast to the lowest supported resolution "timedelta64[s]" (:issue:`49014`)
-- Passing a ``np.datetime64`` object with non-nanosecond resolution to :class:`Timestamp` will retain the input resolution if it is "s", "ms", or "ns"; otherwise it will be cast to the closest supported resolution (:issue:`49008`)
+- Passing a ``np.datetime64`` object with non-nanosecond resolution to :class:`Timestamp` will retain the input resolution if it is "s", "ms", "us", or "ns"; otherwise it will be cast to the closest supported resolution (:issue:`49008`)
+- Passing a string in ISO-8601 format to :class:`Timestamp` will retain the resolution of the parsed input if it is "s", "ms", "us", or "ns"; otherwise it will be cast to the closest supported resolution (:issue:`49737`)
 - The ``other`` argument in :meth:`DataFrame.mask` and :meth:`Series.mask` now defaults to ``no_default`` instead of ``np.nan`` consistent with :meth:`DataFrame.where` and :meth:`Series.where`. Entries will be filled with the corresponding NULL value (``np.nan`` for numpy dtypes, ``pd.NA`` for extension dtypes). (:issue:`49111`)
 - Changed behavior of :meth:`Series.quantile` and :meth:`DataFrame.quantile` with :class:`SparseDtype` to retain sparse dtype (:issue:`49583`)
 - When creating a :class:`Series` with a object-dtype :class:`Index` of datetime objects, pandas no longer silently converts the index to a :class:`DatetimeIndex` (:issue:`39307`, :issue:`23598`)
@@ -798,6 +799,7 @@ Datetimelike
 - Bug in :func:`to_datetime` was raising ``ValueError`` when parsing empty string and non-ISO8601 format was passed. Now, empty strings will be parsed as :class:`NaT`, for compatibility with how is done for ISO8601 formats (:issue:`50251`)
 - Bug in :class:`Timestamp` was showing ``UserWarning``, which was not actionable by users, when parsing non-ISO8601 delimited date strings (:issue:`50232`)
 - Bug in :func:`to_datetime` was showing misleading ``ValueError`` when parsing dates with format containing ISO week directive and ISO weekday directive (:issue:`50308`)
+- Bug in :meth:`Timestamp.round` when the ``freq`` argument has zero-duration (e.g. "0ns") returning incorrect results instead of raising (:issue:`49737`)
 - Bug in :func:`to_datetime` was not raising ``ValueError`` when invalid format was passed and ``errors`` was ``'ignore'`` or ``'coerce'`` (:issue:`50266`)
 - Bug in :class:`DateOffset` was throwing ``TypeError`` when constructing with milliseconds and another super-daily argument (:issue:`49897`)
 -

pandas/_libs/tslibs/conversion.pyx

@@ -405,7 +405,8 @@ cdef _TSObject convert_datetime_to_tsobject(
 
 
 cdef _TSObject _create_tsobject_tz_using_offset(npy_datetimestruct dts,
-                                                int tzoffset, tzinfo tz=None):
+                                                int tzoffset, tzinfo tz=None,
+                                                NPY_DATETIMEUNIT reso=NPY_FR_ns):
     """
     Convert a datetimestruct `dts`, along with initial timezone offset
     `tzoffset` to a _TSObject (with timezone object `tz` - optional).
@@ -416,6 +417,7 @@ cdef _TSObject _create_tsobject_tz_using_offset(npy_datetimestruct dts,
     tzoffset: int
     tz : tzinfo or None
         timezone for the timezone-aware output.
+    reso : NPY_DATETIMEUNIT, default NPY_FR_ns
 
     Returns
     -------
@@ -427,16 +429,19 @@ cdef _TSObject _create_tsobject_tz_using_offset(npy_datetimestruct dts,
         datetime dt
         Py_ssize_t pos
 
-    value = npy_datetimestruct_to_datetime(NPY_FR_ns, &dts)
+    value = npy_datetimestruct_to_datetime(reso, &dts)
     obj.dts = dts
     obj.tzinfo = timezone(timedelta(minutes=tzoffset))
-    obj.value = tz_localize_to_utc_single(value, obj.tzinfo)
+    obj.value = tz_localize_to_utc_single(
+        value, obj.tzinfo, ambiguous=None, nonexistent=None, creso=reso
+    )
+    obj.creso = reso
     if tz is None:
-        check_overflows(obj, NPY_FR_ns)
+        check_overflows(obj, reso)
         return obj
 
     cdef:
-        Localizer info = Localizer(tz, NPY_FR_ns)
+        Localizer info = Localizer(tz, reso)
 
     # Infer fold from offset-adjusted obj.value
     # see PEP 495 https://www.python.org/dev/peps/pep-0495/#the-fold-attribute
@@ -454,6 +459,7 @@ cdef _TSObject _create_tsobject_tz_using_offset(npy_datetimestruct dts,
                   obj.dts.us, obj.tzinfo, fold=obj.fold)
     obj = convert_datetime_to_tsobject(
         dt, tz, nanos=obj.dts.ps // 1000)
+    obj.ensure_reso(reso)  # TODO: more performant to get reso right up front?
     return obj
 
 
@@ -490,7 +496,7 @@ cdef _TSObject _convert_str_to_tsobject(object ts, tzinfo tz, str unit,
         int out_local = 0, out_tzoffset = 0, string_to_dts_failed
         datetime dt
         int64_t ival
-        NPY_DATETIMEUNIT out_bestunit
+        NPY_DATETIMEUNIT out_bestunit, reso
 
     if len(ts) == 0 or ts in nat_strings:
         ts = NaT
@@ -513,19 +519,26 @@ cdef _TSObject _convert_str_to_tsobject(object ts, tzinfo tz, str unit,
             &out_tzoffset, False
         )
         if not string_to_dts_failed:
+            reso = get_supported_reso(out_bestunit)
             try:
-                check_dts_bounds(&dts, NPY_FR_ns)
+                check_dts_bounds(&dts, reso)
                 if out_local == 1:
-                    return _create_tsobject_tz_using_offset(dts,
-                                                            out_tzoffset, tz)
+                    return _create_tsobject_tz_using_offset(
+                        dts, out_tzoffset, tz, reso
+                    )
                 else:
-                    ival = npy_datetimestruct_to_datetime(NPY_FR_ns, &dts)
+                    ival = npy_datetimestruct_to_datetime(reso, &dts)
                     if tz is not None:
                         # shift for _localize_tso
-                        ival = tz_localize_to_utc_single(ival, tz,
-                                                         ambiguous="raise")
-
-                    return convert_to_tsobject(ival, tz, None, False, False)
+                        ival = tz_localize_to_utc_single(
+                            ival, tz, ambiguous="raise", nonexistent=None, creso=reso
+                        )
+                    obj = _TSObject()
+                    obj.dts = dts
+                    obj.value = ival
+                    obj.creso = reso
+                    maybe_localize_tso(obj, tz, obj.creso)
+                    return obj
 
             except OutOfBoundsDatetime:
                 # GH#19382 for just-barely-OutOfBounds falling back to dateutil
@@ -538,10 +551,12 @@ cdef _TSObject _convert_str_to_tsobject(object ts, tzinfo tz, str unit,
                 pass
 
         try:
-            dt = parse_datetime_string(ts, dayfirst=dayfirst,
-                                       yearfirst=yearfirst)
-        except (ValueError, OverflowError):
-            raise ValueError("could not convert string to Timestamp")
+            # TODO: use the one that returns reso
+            dt = parse_datetime_string(
+                ts, dayfirst=dayfirst, yearfirst=yearfirst
+            )
+        except (ValueError, OverflowError) as err:
+            raise ValueError("could not convert string to Timestamp") from err
 
     return convert_datetime_to_tsobject(dt, tz)
 

pandas/_libs/tslibs/offsets.pyx

@@ -162,7 +162,11 @@ def apply_wraps(func):
 
         result = func(self, other)
 
-        result = (<_Timestamp>Timestamp(result))._as_creso(other._creso)
+        result2 = Timestamp(result).as_unit(other.unit)
+        if result == result2:
+            # i.e. the conversion is non-lossy, not the case for e.g.
+            #  test_milliseconds_combination
+            result = result2
 
         if self._adjust_dst:
             result = result.tz_localize(tz)

pandas/_libs/tslibs/timestamps.pyx

@@ -448,6 +448,7 @@ cdef class _Timestamp(ABCTimestamp):
             # cython semantics, args have been switched and this is __radd__
             # TODO(cython3): remove this it moved to __radd__
             return other.__add__(self)
+
         return NotImplemented
 
     def __radd__(self, other):
@@ -1560,8 +1561,17 @@ class Timestamp(_Timestamp):
         cdef:
             int64_t nanos
 
-        to_offset(freq).nanos  # raises on non-fixed freq
-        nanos = delta_to_nanoseconds(to_offset(freq), self._creso)
+        freq = to_offset(freq)
+        freq.nanos  # raises on non-fixed freq
+        nanos = delta_to_nanoseconds(freq, self._creso)
+        if nanos == 0:
+            if freq.nanos == 0:
+                raise ValueError("Division by zero in rounding")
+
+            # e.g. self.unit == "s" and sub-second freq
+            return self
+
+        # TODO: problem if nanos==0
 
         if self.tz is not None:
             value = self.tz_localize(None).value

pandas/core/computation/pytables.py

@@ -11,7 +11,6 @@
 import numpy as np
 
 from pandas._libs.tslibs import (
-    NaT,
     Timedelta,
     Timestamp,
 )
@@ -216,17 +215,16 @@ def stringify(value):
             if isinstance(v, (int, float)):
                 v = stringify(v)
             v = ensure_decoded(v)
-            v = Timestamp(v)
-            if v is not NaT:
-                v = v.as_unit("ns")  # pyright: ignore[reportGeneralTypeIssues]
+            v = Timestamp(v).as_unit("ns")
             if v.tz is not None:
                 v = v.tz_convert("UTC")
             return TermValue(v, v.value, kind)
         elif kind in ("timedelta64", "timedelta"):
             if isinstance(v, str):
-                v = Timedelta(v).value
+                v = Timedelta(v)
             else:
-                v = Timedelta(v, unit="s").value
+                v = Timedelta(v, unit="s")
+            v = v.as_unit("ns").value
             return TermValue(int(v), v, kind)
         elif meta == "category":
             metadata = extract_array(self.metadata, extract_numpy=True)

pandas/core/resample.py

@@ -2085,7 +2085,7 @@ def _adjust_dates_anchored(
     elif origin == "start":
         origin_nanos = first.value
     elif isinstance(origin, Timestamp):
-        origin_nanos = origin.value
+        origin_nanos = origin.as_unit("ns").value
     elif origin in ["end", "end_day"]:
         origin_last = last if origin == "end" else last.ceil("D")
         sub_freq_times = (origin_last.value - first.value) // freq.nanos

pandas/tests/arithmetic/test_datetime64.py

@@ -1699,15 +1699,15 @@ def test_datetimeindex_sub_timestamp_overflow(self):
         dtimax = pd.to_datetime(["2021-12-28 17:19", Timestamp.max])
         dtimin = pd.to_datetime(["2021-12-28 17:19", Timestamp.min])
 
-        tsneg = Timestamp("1950-01-01")
+        tsneg = Timestamp("1950-01-01").as_unit("ns")
         ts_neg_variants = [
             tsneg,
             tsneg.to_pydatetime(),
             tsneg.to_datetime64().astype("datetime64[ns]"),
             tsneg.to_datetime64().astype("datetime64[D]"),
         ]
 
-        tspos = Timestamp("1980-01-01")
+        tspos = Timestamp("1980-01-01").as_unit("ns")
         ts_pos_variants = [
             tspos,
             tspos.to_pydatetime(),

pandas/tests/arrays/test_timedeltas.py

@@ -102,7 +102,7 @@ def test_add_pdnat(self, tda):
     # TODO: 2022-07-11 this is the only test that gets to DTA.tz_convert
     #  or tz_localize with non-nano; implement tests specific to that.
     def test_add_datetimelike_scalar(self, tda, tz_naive_fixture):
-        ts = pd.Timestamp("2016-01-01", tz=tz_naive_fixture)
+        ts = pd.Timestamp("2016-01-01", tz=tz_naive_fixture).as_unit("ns")
 
         expected = tda.as_unit("ns") + ts
         res = tda + ts

pandas/tests/indexes/datetimes/methods/test_astype.py

@@ -276,7 +276,7 @@ def _check_rng(rng):
     )
     def test_integer_index_astype_datetime(self, tz, dtype):
         # GH 20997, 20964, 24559
-        val = [Timestamp("2018-01-01", tz=tz).value]
+        val = [Timestamp("2018-01-01", tz=tz).as_unit("ns").value]
         result = Index(val, name="idx").astype(dtype)
         expected = DatetimeIndex(["2018-01-01"], tz=tz, name="idx")
         tm.assert_index_equal(result, expected)

pandas/tests/indexes/datetimes/test_constructors.py

@@ -804,7 +804,7 @@ def test_constructor_timestamp_near_dst(self):
     )
     def test_constructor_with_int_tz(self, klass, box, tz, dtype):
         # GH 20997, 20964
-        ts = Timestamp("2018-01-01", tz=tz)
+        ts = Timestamp("2018-01-01", tz=tz).as_unit("ns")
         result = klass(box([ts.value]), dtype=dtype)
         expected = klass([ts])
         assert result == expected

pandas/tests/io/json/test_pandas.py

@@ -973,7 +973,9 @@ def test_mixed_timedelta_datetime(self):
         ts = Timestamp("20130101")
         frame = DataFrame({"a": [td, ts]}, dtype=object)
 
-        expected = DataFrame({"a": [pd.Timedelta(td).as_unit("ns").value, ts.value]})
+        expected = DataFrame(
+            {"a": [pd.Timedelta(td).as_unit("ns").value, ts.as_unit("ns").value]}
+        )
         result = read_json(frame.to_json(date_unit="ns"), dtype={"a": "int64"})
         tm.assert_frame_equal(result, expected, check_index_type=False)
 

pandas/tests/scalar/timedelta/test_arithmetic.py

@@ -99,13 +99,14 @@ def test_td_add_datetimelike_scalar(self, op):
         assert result is NaT
 
     def test_td_add_timestamp_overflow(self):
+        ts = Timestamp("1700-01-01").as_unit("ns")
         msg = "Cannot cast 259987 from D to 'ns' without overflow."
         with pytest.raises(OutOfBoundsTimedelta, match=msg):
-            Timestamp("1700-01-01") + Timedelta(13 * 19999, unit="D")
+            ts + Timedelta(13 * 19999, unit="D")
 
         msg = "Cannot cast 259987 days 00:00:00 to unit='ns' without overflow"
         with pytest.raises(OutOfBoundsTimedelta, match=msg):
-            Timestamp("1700-01-01") + timedelta(days=13 * 19999)
+            ts + timedelta(days=13 * 19999)
 
     @pytest.mark.parametrize("op", [operator.add, ops.radd])
     def test_td_add_td(self, op):

pandas/tests/scalar/timestamp/test_arithmetic.py

@@ -38,7 +38,7 @@ def test_overflow_offset_raises(self):
         # xref https://github.com/statsmodels/statsmodels/issues/3374
         # ends up multiplying really large numbers which overflow
 
-        stamp = Timestamp("2017-01-13 00:00:00")
+        stamp = Timestamp("2017-01-13 00:00:00").as_unit("ns")
         offset_overflow = 20169940 * offsets.Day(1)
         msg = (
             "the add operation between "
@@ -59,7 +59,7 @@ def test_overflow_offset_raises(self):
         # xref https://github.com/pandas-dev/pandas/issues/14080
         # used to crash, so check for proper overflow exception
 
-        stamp = Timestamp("2000/1/1")
+        stamp = Timestamp("2000/1/1").as_unit("ns")
         offset_overflow = to_offset("D") * 100**5
 
         lmsg3 = (
@@ -77,8 +77,8 @@ def test_overflow_offset_raises(self):
     def test_overflow_timestamp_raises(self):
         # https://github.com/pandas-dev/pandas/issues/31774
         msg = "Result is too large"
-        a = Timestamp("2101-01-01 00:00:00")
-        b = Timestamp("1688-01-01 00:00:00")
+        a = Timestamp("2101-01-01 00:00:00").as_unit("ns")
+        b = Timestamp("1688-01-01 00:00:00").as_unit("ns")
 
         with pytest.raises(OutOfBoundsDatetime, match=msg):
             a - b
@@ -239,7 +239,7 @@ def test_add_int_with_freq(self, ts, other):
     @pytest.mark.parametrize("shape", [(6,), (2, 3)])
     def test_addsub_m8ndarray(self, shape):
         # GH#33296
-        ts = Timestamp("2020-04-04 15:45")
+        ts = Timestamp("2020-04-04 15:45").as_unit("ns")
         other = np.arange(6).astype("m8[h]").reshape(shape)
 
         result = ts + other