flox: don't set fill_value where possible

doc/whats-new.rst

@@ -60,6 +60,8 @@ Bug fixes
 - Fix deprecation warning that was raised when calling ``np.array`` on an ``xr.DataArray``
   in NumPy 2.0 (:issue:`9312`, :pull:`9393`)
   By `Andrew Scherer <https://github.com/andrew-s28>`_.
+- Fix a few bugs affecting groupby reductions with `flox`. (:issue:`8090`, :issue:`9398`).
+  By `Deepak Cherian <https://github.com/dcherian>`_.
 
 Performance
 ~~~~~~~~~~~

xarray/core/dataarray.py

@@ -6785,8 +6785,8 @@ def groupby(
 
         >>> da.groupby("letters").sum()
         <xarray.DataArray (letters: 2, y: 3)> Size: 48B
-        array([[ 9., 11., 13.],
-               [ 9., 11., 13.]])
+        array([[ 9, 11, 13],
+               [ 9, 11, 13]])
         Coordinates:
           * letters  (letters) object 16B 'a' 'b'
         Dimensions without coordinates: y

xarray/core/dataset.py

@@ -10387,7 +10387,7 @@ def groupby(
           * letters  (letters) object 16B 'a' 'b'
         Dimensions without coordinates: y
         Data variables:
-            foo      (letters, y) float64 48B 9.0 11.0 13.0 9.0 11.0 13.0
+            foo      (letters, y) int64 48B 9 11 13 9 11 13
 
         Grouping by multiple variables
 

xarray/core/groupby.py

@@ -787,14 +787,12 @@ def _maybe_restore_empty_groups(self, combined):
         """Our index contained empty groups (e.g., from a resampling or binning). If we
         reduced on that dimension, we want to restore the full index.
         """
-        from xarray.groupers import BinGrouper, TimeResampler
-
+        has_missing_groups = (
+            self.encoded.unique_coord.size != self.encoded.full_index.size
+        )
         indexers = {}
         for grouper in self.groupers:
-            if (
-                isinstance(grouper.grouper, BinGrouper | TimeResampler)
-                and grouper.name in combined.dims
-            ):
+            if has_missing_groups and grouper.name in combined._indexes:
                 indexers[grouper.name] = grouper.full_index
         if indexers:
             combined = combined.reindex(**indexers)
@@ -849,10 +847,6 @@ def _flox_reduce(
             else obj._coords
         )
 
-        any_isbin = any(
-            isinstance(grouper.grouper, BinGrouper) for grouper in self.groupers
-        )
-
         if keep_attrs is None:
             keep_attrs = _get_keep_attrs(default=True)
 
@@ -926,14 +920,14 @@ def _flox_reduce(
             ):
                 raise ValueError(f"cannot reduce over dimensions {dim}.")
 
-        if kwargs["func"] not in ["all", "any", "count"]:
-            kwargs.setdefault("fill_value", np.nan)
-        if any_isbin and kwargs["func"] == "count":
-            # This is an annoying hack. Xarray returns np.nan
-            # when there are no observations in a bin, instead of 0.
-            # We can fake that here by forcing min_count=1.
-            # note min_count makes no sense in the xarray world
-            # as a kwarg for count, so this should be OK
+        has_missing_groups = (
+            self.encoded.unique_coord.size != self.encoded.full_index.size
+        )
+        if has_missing_groups or kwargs.get("min_count", 0) > 0:
+            # Xarray *always* returns np.nan when there are no observations in a group,
+            # We can fake that here by forcing min_count=1 when it is not set.
+            # This handles boolean reductions, and count
+            # See GH8090, GH9398
             kwargs.setdefault("fill_value", np.nan)
             kwargs.setdefault("min_count", 1)
 

xarray/tests/test_groupby.py

@@ -2784,6 +2784,58 @@
     # ------
 
 
+@pytest.mark.parametrize(
+    "reduction", ["max", "min", "nanmax", "nanmin", "sum", "nansum", "prod", "nanprod"]
+)
+def test_groupby_preserve_dtype(reduction):
+    # all groups are present, we should follow numpy exactly
+    ds = xr.Dataset(
+        {
+            "test": (
+                ["x", "y"],
+                np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype="int16"),
+            )
+        },
+        coords={"idx": ("x", [1, 2, 1])},
+    )
+
+    kwargs = {}
+    if "nan" in reduction:
+        kwargs["skipna"] = True
+    # TODO: fix dtype with numbagg/bottleneck and use_flox=False
+    with xr.set_options(use_numbagg=False, use_bottleneck=False):
+        actual = getattr(ds.groupby("idx"), reduction.removeprefix("nan"))(
+            **kwargs
+        ).test.dtype
+    expected = getattr(np, reduction)(ds.test.data, axis=0).dtype
+
+    assert actual == expected
+
+
+@requires_flox
+@pytest.mark.parametrize("reduction", ["any", "all", "count"])
+def test_gappy_resample_reductions(reduction):
+    # GH8090
+    dates = (("1988-12-01", "1990-11-30"), ("2000-12-01", "2001-11-30"))
+    times = [xr.date_range(*d, freq="D") for d in dates]
+
+    da = xr.concat(
+        [
+            xr.DataArray(np.random.rand(len(t)), coords={"time": t}, dims="time")
+            for t in times
+        ],
+        dim="time",
+    ).chunk(time=100)
+
+    rs = (da > 0.5).resample(time="YS-DEC")
+    method = getattr(rs, reduction)
+    with xr.set_options(use_flox=True):
+        actual = method(dim="time")
+    with xr.set_options(use_flox=False):
+        expected = method(dim="time")
+    assert_identical(expected, actual)
+
+
 # Possible property tests
 # 1. lambda x: x
 # 2. grouped-reduce on unique coords is identical to array