FIX: prevent unnecessary type casts in rolling mean (#10341)

* FIX: use dtype from intermediate sum for casting of count instead of current "int" when calculating mean in rolling operations to prevent unnecessary casting to/from int

* Add code comment and whats-new.rst entry

Author: kmuehlbauer

doc/whats-new.rst

@@ -45,6 +45,7 @@ Bug fixes
 
 - Allow accessing arbitrary attributes on Pandas ExtensionArrays.
   By `Deepak Cherian <https://github.com/dcherian>`_.
+- Use dtype from intermediate sum instead of source dtype or "int" for casting of count when calculating mean in rolling for correct operations (preserve float dtypes, correct mean of bool arrays) (:issue:`10340`, :pull:`10341`).
 
 
 Documentation

xarray/computation/rolling.py

@@ -195,14 +195,15 @@ def method(self, keep_attrs=None, **kwargs):
         return method
 
     def _mean(self, keep_attrs, **kwargs):
-        result = self.sum(keep_attrs=False, **kwargs) / duck_array_ops.astype(
-            self.count(keep_attrs=False), dtype=int, copy=False
+        result = self.sum(keep_attrs=False, **kwargs)
+        # use dtype of result for casting of count
+        # this allows for GH #7062 and GH #8864, fixes GH #10340
+        result /= duck_array_ops.astype(
+            self.count(keep_attrs=False), dtype=result.dtype, copy=False
         )
         if keep_attrs:
             result.attrs = self.obj.attrs
 
-        if self.obj.dtype.kind not in "bi":
-            result = result.astype(self.obj.dtype, copy=False)
         return result
 
     _mean.__doc__ = _ROLLING_REDUCE_DOCSTRING_TEMPLATE.format(name="mean")