PERF: groupby iteration

pandas/core/generic.py

@@ -469,6 +469,32 @@ def _validate_dtype(cls, dtype) -> DtypeObj | None:
 
         return dtype
 
+    @property
+    def _can_fast_construct(self) -> bool_t:
+        """
+        Check if we can avoid _constructor lookup and __finalize__ calls.
+        """
+        # GH#46505 repeated __finalize__ calls caused perf regression,
+        #  see if we can avoid those in some cases.
+        from pandas import (
+            DataFrame,
+            Series,
+        )
+
+        if self.attrs:
+            # __finalize__ is not a no-op
+            return False
+
+        if type(self) is Series:
+            if self._metadata == ["name"]:
+                return True
+            return False
+        elif type(self) is DataFrame:
+            return not self._metadata
+
+        # subclass, may have overridden _constructor -> cannot fast-construct
+        return False
+
     # ----------------------------------------------------------------------
     # Construction
 

pandas/core/groupby/ops.py

@@ -1257,8 +1257,13 @@ def __iter__(self) -> Iterator:
 
         starts, ends = lib.generate_slices(self._slabels, self.ngroups)
 
+        if sdata._can_fast_construct:
+            chop = self._chop_fast
+        else:
+            chop = self._chop
+
         for start, end in zip(starts, ends):
-            yield self._chop(sdata, slice(start, end))
+            yield chop(sdata, slice(start, end))
 
     @cache_readonly
     def _sorted_data(self) -> NDFrameT:
@@ -1267,6 +1272,9 @@ def _sorted_data(self) -> NDFrameT:
     def _chop(self, sdata, slice_obj: slice) -> NDFrame:
         raise AbstractMethodError(self)
 
+    def _chop_fast(self, sdata, slice_obj: slice) -> NDFrame:
+        raise AbstractMethodError(self)
+
 
 class SeriesSplitter(DataSplitter):
     def _chop(self, sdata: Series, slice_obj: slice) -> Series:
@@ -1275,6 +1283,12 @@ def _chop(self, sdata: Series, slice_obj: slice) -> Series:
         ser = sdata._constructor(mgr, name=sdata.name, fastpath=True)
         return ser.__finalize__(sdata, method="groupby")
 
+    def _chop_fast(self, sdata: Series, slice_obj: slice) -> Series:
+        # _chop specialized to cast with _can_fast_construct
+        mgr = sdata._mgr.get_slice(slice_obj)
+        ser = Series(mgr, name=sdata.name, fastpath=True)
+        return ser
+
 
 class FrameSplitter(DataSplitter):
     def _chop(self, sdata: DataFrame, slice_obj: slice) -> DataFrame:
@@ -1287,6 +1301,12 @@ def _chop(self, sdata: DataFrame, slice_obj: slice) -> DataFrame:
         df = sdata._constructor(mgr)
         return df.__finalize__(sdata, method="groupby")
 
+    def _chop_fast(self, sdata: DataFrame, slice_obj: slice) -> DataFrame:
+        # _chop specialized to cast with _can_fast_construct
+        mgr = sdata._mgr.get_slice(slice_obj, axis=1 - self.axis)
+        df = DataFrame(mgr)
+        return df
+
 
 def get_splitter(
     data: NDFrame, labels: np.ndarray, ngroups: int, axis: AxisInt = 0