fix(dtypes): dayfirst=True no longer corrupts ISO-8601 dates

CHANGELOG.md

@@ -6,6 +6,34 @@ adheres to [Semantic Versioning](https://semver.org/).
 
 ## [Unreleased]
 
+### Fixed
+
+- `dayfirst=True` no longer reinterprets unambiguous ISO-8601 dates.
+  `fd.clean(df, dayfirst=True)` on `["2021-01-05", "2021-02-11"]` silently
+  returned `2021-05-01`, `2021-11-02` — month and day swapped, with no warning
+  and no coercion record. `dayfirst` resolves *ambiguous* short dates such as
+  `05/12/2021`; an ISO-8601 date is `YYYY-MM-DD` by definition and has no
+  ambiguity to resolve. Cause: pandas 2 infers one format for a whole column
+  from its first value and, under `dayfirst=True`, reads an ISO date as
+  `%Y-%d-%m`; `_parse_datetime` now passes `format="mixed"` in that case so
+  each value is read by its own shape. pandas 1.x infers per value and was
+  never affected, and has no `format="mixed"`, so the change is guarded on the
+  pandas major version.
+
+  The corruption was data-dependent, which is what made it easy to miss: it was
+  silent only while every day was `<= 12`, because a day `>= 13` made the
+  guessed format fail, dropped the parse share below `datetime_threshold`, and
+  triggered the mixed-format retry that produced the correct reading. The same
+  column therefore read correctly or incorrectly depending on values it
+  happened to contain, or on an unrelated threshold.
+
+  **Compatibility impact:** output changes for ISO-8601 columns cleaned with
+  `dayfirst=True` on pandas 2 — from a wrong reading to the correct one.
+  `dayfirst` behaviour on genuinely ambiguous slash dates is unchanged, as are
+  the `dayfirst="auto"` and `semantic_context` routes, which were never
+  affected.
+
+
 ### Documentation
 
 - `StreamingCleanConfig.window_size` was documented as sizing "rolling

src/freshdata/steps/dtypes.py

@@ -381,7 +381,16 @@ def _parse_datetime(
     s: pd.Series, mixed_formats: bool, dayfirst: bool = False
 ) -> pd.Series | None:
     kwargs: dict = {"errors": "coerce", "dayfirst": dayfirst}
-    if mixed_formats:
+    if mixed_formats or (dayfirst and PANDAS_MAJOR >= 2):
+        # ``dayfirst`` exists to resolve *ambiguous* short dates like
+        # ``05/12/2021``. An ISO-8601 date is not ambiguous. But pandas 2's
+        # format inference infers a single format for the whole column from
+        # the first value, and with ``dayfirst=True`` it reads ``2021-01-05``
+        # as ``%Y-%d-%m`` -- silently returning 2021-05-01. ``format="mixed"``
+        # parses each value by its own apparent format, which fixes ISO input
+        # while leaving genuinely ambiguous slash dates to ``dayfirst``.
+        # pandas 1.x infers per value already and is unaffected; it also has
+        # no ``format="mixed"``, hence the version guard.
         kwargs["format"] = "mixed"
     with warnings.catch_warnings():
         warnings.simplefilter("ignore")  # format-inference chatter; report covers it

tests/test_dayfirst_iso_dates.py

@@ -0,0 +1,81 @@
+"""``dayfirst=True`` must not reinterpret unambiguous ISO-8601 dates.
+
+``dayfirst`` exists to resolve *ambiguous* short dates such as ``05/12/2021``.
+An ISO-8601 date is ``YYYY-MM-DD`` by definition, so there is nothing for it to
+resolve. Before this fix ``fd.clean(df, dayfirst=True)`` silently returned
+``2021-05-01`` for the input ``2021-01-05``.
+
+The cause was pandas 2's format inference, which picks ONE format for a whole
+column from its first value and, with ``dayfirst=True``, reads an ISO date as
+``%Y-%d-%m``. pandas 1.x infers per value and was never affected.
+
+The corruption was data-dependent, which is what made it dangerous: it is
+silent only while every day is <= 12. A day >= 13 makes the guessed format fail
+on that value, dropping the parse share below ``datetime_threshold`` so the
+mixed-format retry replaces the column with the correct reading. The same
+column therefore read correctly or incorrectly depending on values it happened
+to contain, or on an unrelated threshold.
+"""
+
+from __future__ import annotations
+
+import pandas as pd
+import pytest
+
+import freshdata as fd
+
+ISO = ["2021-01-05", "2021-02-11", "2021-03-09"]
+
+
+def _dates(frame, column="v"):
+    return [str(v.date()) for v in frame[column]]
+
+
+def test_iso_dates_are_not_reinterpreted_when_dayfirst_is_true():
+    """The regression. Before the fix this returned 2021-05-01 and friends."""
+    out = fd.clean(pd.DataFrame({"v": ISO}), dayfirst=True, verbose=False)
+    assert _dates(out) == ISO
+
+
+@pytest.mark.parametrize("dayfirst", [True, False, "auto", None])
+def test_iso_dates_read_the_same_whatever_dayfirst_says(dayfirst):
+    """``dayfirst`` is not a question ISO-8601 input can answer differently."""
+    kwargs = {} if dayfirst is None else {"dayfirst": dayfirst}
+    out = fd.clean(pd.DataFrame({"v": ISO}), verbose=False, **kwargs)
+    assert _dates(out) == ISO
+
+
+def test_dayfirst_still_does_its_actual_job_on_ambiguous_slash_dates():
+    """The fix must not disarm ``dayfirst`` where it is genuinely needed."""
+    ambiguous = ["05/12/2021", "06/11/2021"]
+    day_first = fd.clean(pd.DataFrame({"v": ambiguous}), dayfirst=True, verbose=False)
+    assert _dates(day_first) == ["2021-12-05", "2021-11-06"]
+
+    month_first = fd.clean(pd.DataFrame({"v": ambiguous}), dayfirst=False, verbose=False)
+    assert _dates(month_first) == ["2021-05-12", "2021-06-11"]
+
+
+def test_the_reading_no_longer_depends_on_whether_a_day_exceeds_twelve():
+    """The data-dependence that made the corruption hard to notice.
+
+    Nineteen consecutive January dates. Before the fix these read correctly at
+    the default threshold but became 2021-01-01, 2021-02-01, 2021-03-01 ... at
+    ``datetime_threshold=0.5``, losing seven values to ``NaT``.
+    """
+    dates = [f"2021-01-{d:02d}" for d in range(1, 20)]
+
+    default = fd.clean(pd.DataFrame({"v": dates}), dayfirst=True, verbose=False)
+    lowered = fd.clean(
+        pd.DataFrame({"v": dates}), dayfirst=True, datetime_threshold=0.5, verbose=False
+    )
+
+    assert _dates(default) == dates
+    assert _dates(lowered) == dates
+    assert not default["v"].isna().any()
+    assert not lowered["v"].isna().any()
+
+
+def test_a_column_mixing_iso_and_ambiguous_forms_reads_each_by_its_own_shape():
+    mixed = ["2021-01-05", "06/11/2021", "2021-03-09"]
+    out = fd.clean(pd.DataFrame({"v": mixed}), dayfirst=True, verbose=False)
+    assert _dates(out) == ["2021-01-05", "2021-11-06", "2021-03-09"]