promote floating-point numeric datetimes to 64-bit before decoding

doc/whats-new.rst

@@ -37,6 +37,8 @@ Bug fixes
 ~~~~~~~~~
 - Make :py:func:`testing.assert_allclose` work with numpy 2.0 (:issue:`9165`, :pull:`9166`).
   By `Pontus Lurcock <https://github.com/pont-us>`_.
+- Promote floating-point numeric datetimes before decoding (:issue:`9179`, :pull:`9182`).
+  By `Justus Magin <https://github.com/keewis>`_.
 
 
 Documentation

xarray/coding/times.py

@@ -278,6 +278,8 @@ def _decode_datetime_with_pandas(
     # timedelta64 value, and therefore would raise an error in the lines above.
     if flat_num_dates.dtype.kind in "iu":
         flat_num_dates = flat_num_dates.astype(np.int64)
+    elif flat_num_dates.dtype.kind in "f":
+        flat_num_dates = flat_num_dates.astype(np.float64)
 
     # Cast input ordinals to integers of nanoseconds because pd.to_timedelta
     # works much faster when dealing with integers (GH 1399).

xarray/tests/test_coding_times.py

@@ -1182,6 +1182,22 @@ def test_decode_0size_datetime(use_cftime):
     np.testing.assert_equal(expected, actual)
 
 
+def test_decode_float_datetime():
+    num_dates = np.array([1867128, 1867134, 1867140], dtype="float32")
+    units = "hours since 1800-01-01"
+    calendar = "standard"
+
+    expected = np.array(
+        ["2013-01-01T00:00:00", "2013-01-01T06:00:00", "2013-01-01T12:00:00"],
+        dtype="datetime64[ns]",
+    )
+
+    actual = decode_cf_datetime(
+        num_dates, units=units, calendar=calendar, use_cftime=False
+    )
+    np.testing.assert_equal(actual, expected)
+
+
 @requires_cftime
 def test_scalar_unit() -> None:
     # test that a scalar units (often NaN when using to_netcdf) does not raise an error