BUG: Fix near-minimum timestamp handling

doc/source/whatsnew/v2.2.1.rst

@@ -15,6 +15,7 @@ Fixed regressions
 ~~~~~~~~~~~~~~~~~
 - Fixed memory leak in :func:`read_csv` (:issue:`57039`)
 - Fixed performance regression in :meth:`Series.combine_first` (:issue:`55845`)
+- Fixed regression causing overflow for near-minimum timestamps (:issue:`57150`)
 - Fixed regression in :func:`concat` changing long-standing behavior that always sorted the non-concatenation axis when the axis was a :class:`DatetimeIndex` (:issue:`57006`)
 - Fixed regression in :func:`merge_ordered` raising ``TypeError`` for ``fill_method="ffill"`` and ``how="left"`` (:issue:`57010`)
 - Fixed regression in :func:`pandas.testing.assert_series_equal` defaulting to ``check_exact=True`` when checking the :class:`Index` (:issue:`57067`)

pandas/_libs/src/vendored/numpy/datetime/np_datetime.c

@@ -482,10 +482,33 @@ npy_datetime npy_datetimestruct_to_datetime(NPY_DATETIMEUNIT base,
 
   if (base == NPY_FR_ns) {
     int64_t nanoseconds;
-    PD_CHECK_OVERFLOW(
-        scaleMicrosecondsToNanoseconds(microseconds, &nanoseconds));
-    PD_CHECK_OVERFLOW(
-        checked_int64_add(nanoseconds, dts->ps / 1000, &nanoseconds));
+
+    // Minimum valid timestamp in nanoseconds (1677-09-21 00:12:43.145224193).
+    const int64_t min_nanoseconds = NPY_MIN_INT64 + 1;
+
+    // For near-minimum timestamps (1677-09-21 00:12:43.145224193 through
+    // 1677-09-21 00:12:43.145224999), scaling microseconds to nanoseconds
+    // overflows (1677-09-21 00:12:43.145224 -> 1677-09-21 00:12:43.145224000),
+    // but adding nanoseconds can put the timestamp back in a valid range for
+    // nanosecond parts >= 193.
+
+    // (min_nanoseconds / 1000 - 1) * 1000 would overflow, so do not scale.
+    // This happens if microseconds corresponds to 1677-09-21 00:12:43.145224.
+    if (microseconds == min_nanoseconds / 1000 - 1) {
+      // Instead, use minimum nanosecond timestamp as base and offset it with
+      // nanosecond delta between dts and the minimum (_NS_MIN_DTS.ps = 193000).
+      // If dts->ps >= _NS_MIN_DTS.ps, timestamp is at/above the minimum.
+      // If dts->ps < _NS_MIN_DTS.ps, timestamp is below minimum and overflows.
+      PD_CHECK_OVERFLOW(checked_int64_add(
+          min_nanoseconds, (dts->ps - _NS_MIN_DTS.ps) / 1000, &nanoseconds));
+    } else {
+      // microseconds does not correspond to near-minimum timestamp, use default
+      // scaling and addition approach, handling any other overflows.
+      PD_CHECK_OVERFLOW(
+          scaleMicrosecondsToNanoseconds(microseconds, &nanoseconds));
+      PD_CHECK_OVERFLOW(
+          checked_int64_add(nanoseconds, dts->ps / 1000, &nanoseconds));
+    }
 
     return nanoseconds;
   }

pandas/tests/tslibs/test_array_to_datetime.py

@@ -262,6 +262,23 @@ def test_to_datetime_barely_out_of_bounds():
         tslib.array_to_datetime(arr)
 
 
+@pytest.mark.parametrize(
+    "timestamp",
+    [
+        # Close enough to bounds that scaling micros to nanos overflows
+        # but adding nanos would result in an in-bounds datetime.
+        "1677-09-21T00:12:43.145224193",
+        "1677-09-21T00:12:43.145224999",
+        # this always worked
+        "1677-09-21T00:12:43.145225000",
+    ],
+)
+def test_to_datetime_barely_inside_bounds(timestamp):
+    # see gh-57150
+    result, _ = tslib.array_to_datetime(np.array([timestamp], dtype=object))
+    tm.assert_numpy_array_equal(result, np.array([timestamp], dtype="M8[ns]"))
+
+
 class SubDatetime(datetime):
     pass