From 367f8a18bd04bb9c521ee5788b5fce24be63275d Mon Sep 17 00:00:00 2001
From: jbrockmendel <jbrockmendel@gmail.com>
Date: Mon, 28 Feb 2022 13:44:26 -0800
Subject: [PATCH] REF: implement _reconstruct_ea_result, _ea_to_cython_values
 (#46172)

---
 pandas/_libs/groupby.pyx   |  8 +++++++
 pandas/core/groupby/ops.py | 47 +++++++++++++++++++++++---------------
 2 files changed, 36 insertions(+), 19 deletions(-)

diff --git a/pandas/_libs/groupby.pyx b/pandas/_libs/groupby.pyx
index e9d6b6813c6fd..79457bcd6506f 100644
--- a/pandas/_libs/groupby.pyx
+++ b/pandas/_libs/groupby.pyx
@@ -1350,6 +1350,10 @@ cdef group_min_max(iu_64_floating_t[:, ::1] out,
                     else:
                         if uses_mask:
                             result_mask[i, j] = True
+                            # set out[i, j] to 0 to be deterministic, as
+                            #  it was initialized with np.empty. Also ensures
+                            #  we can downcast out if appropriate.
+                            out[i, j] = 0
                         else:
                             out[i, j] = nan_val
                 else:
@@ -1494,6 +1498,10 @@ cdef group_cummin_max(iu_64_floating_t[:, ::1] out,
                 if not skipna and na_possible and seen_na[lab, j]:
                     if uses_mask:
                         mask[i, j] = 1   # FIXME: shouldn't alter inplace
+                        # Set to 0 ensures that we are deterministic and can
+                        #  downcast if appropriate
+                        out[i, j] = 0
+
                     else:
                         out[i, j] = na_val
                 else:
diff --git a/pandas/core/groupby/ops.py b/pandas/core/groupby/ops.py
index b4fef6fcdff19..8e38d542349d9 100644
--- a/pandas/core/groupby/ops.py
+++ b/pandas/core/groupby/ops.py
@@ -339,6 +339,26 @@ def _ea_wrap_cython_operation(
                 **kwargs,
             )
 
+        npvalues = self._ea_to_cython_values(values)
+
+        res_values = self._cython_op_ndim_compat(
+            npvalues,
+            min_count=min_count,
+            ngroups=ngroups,
+            comp_ids=comp_ids,
+            mask=None,
+            **kwargs,
+        )
+
+        if self.how in ["rank"]:
+            # i.e. how in WrappedCythonOp.cast_blocklist, since
+            #  other cast_blocklist methods dont go through cython_operation
+            return res_values
+
+        return self._reconstruct_ea_result(values, res_values)
+
+    def _ea_to_cython_values(self, values: ExtensionArray):
+        # GH#43682
         if isinstance(values, (DatetimeArray, PeriodArray, TimedeltaArray)):
             # All of the functions implemented here are ordinal, so we can
             #  operate on the tz-naive equivalents
@@ -356,22 +376,7 @@ def _ea_wrap_cython_operation(
             raise NotImplementedError(
                 f"function is not implemented for this dtype: {values.dtype}"
             )
-
-        res_values = self._cython_op_ndim_compat(
-            npvalues,
-            min_count=min_count,
-            ngroups=ngroups,
-            comp_ids=comp_ids,
-            mask=None,
-            **kwargs,
-        )
-
-        if self.how in ["rank"]:
-            # i.e. how in WrappedCythonOp.cast_blocklist, since
-            #  other cast_blocklist methods dont go through cython_operation
-            return res_values
-
-        return self._reconstruct_ea_result(values, res_values)
+        return npvalues
 
     def _reconstruct_ea_result(self, values, res_values):
         """
@@ -387,6 +392,7 @@ def _reconstruct_ea_result(self, values, res_values):
             return cls._from_sequence(res_values, dtype=dtype)
 
         elif needs_i8_conversion(values.dtype):
+            assert res_values.dtype.kind != "f"  # just to be on the safe side
             i8values = res_values.view("i8")
             return type(values)(i8values, dtype=values.dtype)
 
@@ -577,9 +583,12 @@ def _call_cython_op(
                 cutoff = max(1, min_count)
                 empty_groups = counts < cutoff
                 if empty_groups.any():
-                    # Note: this conversion could be lossy, see GH#40767
-                    result = result.astype("float64")
-                    result[empty_groups] = np.nan
+                    if result_mask is not None and self.uses_mask():
+                        assert result_mask[empty_groups].all()
+                    else:
+                        # Note: this conversion could be lossy, see GH#40767
+                        result = result.astype("float64")
+                        result[empty_groups] = np.nan
 
         result = result.T