From cebc07fed0707457487656836e627a154416d252 Mon Sep 17 00:00:00 2001
From: Palash Lalwani <palashlalwani.r@gmail.com>
Date: Wed, 22 Jul 2026 21:24:52 +0530
Subject: [PATCH] fix: don't pass grouping columns to
 groupby().parallel_apply() on pandas >= 3.0

Since pandas 3.0, DataFrameGroupBy.apply operates on _obj_with_exclusions, so the
grouping columns are no longer passed to the applied function. Iterating a
DataFrameGroupBy still yields them, which is how pandarallel builds its chunks, so
parallel_apply kept passing them.

This produced a silent divergence rather than an error: user functions that iterate
columns or aggregate across the frame got different results from parallel_apply than
from apply.

Restrict each group to the columns apply would have used, gated on the pandas version
so behaviour on pandas < 3.0 is unchanged.

The groupby test fixture read df.b, but the test also covers groupby(["a", "b"]) where
b is now a grouping column, making the reference apply() call raise on pandas 3.0.
Switched the fixture to column c, which is never a grouping column in these tests.
---
 pandarallel/data_types/dataframe_groupby.py | 18 +++++++++++++++++-
 tests/test_pandarallel.py                   |  8 ++++++--
 2 files changed, 23 insertions(+), 3 deletions(-)

diff --git a/pandarallel/data_types/dataframe_groupby.py b/pandarallel/data_types/dataframe_groupby.py
index f2dbb91..9b448d7 100644
--- a/pandarallel/data_types/dataframe_groupby.py
+++ b/pandarallel/data_types/dataframe_groupby.py
@@ -17,8 +17,24 @@ def get_chunks(
             chunks = chunk(dataframe_groupby.ngroups, nb_workers)
             iterator = iter(dataframe_groupby)
 
+            # Since pandas 3.0, `DataFrameGroupBy.apply` operates on
+            # `_obj_with_exclusions`, i.e. the grouping columns are no longer passed to
+            # the applied function. Iterating a `DataFrameGroupBy` still yields them
+            # though, so restrict each group to the columns `apply` would have used.
+            # On pandas < 3.0 the grouping columns are still passed, so keep them.
+            columns = (
+                dataframe_groupby._obj_with_exclusions.columns
+                if get_pandas_version() >= (3, 0)
+                else None
+            )
+
             for chunk_ in chunks:
-                yield [next(iterator) for _ in range(chunk_.stop - chunk_.start)]
+                groups = [next(iterator) for _ in range(chunk_.stop - chunk_.start)]
+
+                if columns is not None:
+                    groups = [(key, df[columns]) for key, df in groups]
+
+                yield groups
 
         @staticmethod
         def work(
diff --git a/tests/test_pandarallel.py b/tests/test_pandarallel.py
index 0f91c32..2f43458 100644
--- a/tests/test_pandarallel.py
+++ b/tests/test_pandarallel.py
@@ -93,12 +93,16 @@ def func(x):
 
 @pytest.fixture()
 def func_dataframe_groupby_apply():
+    # Uses column `c`, which is never a grouping column in the tests below. Since pandas
+    # 3.0 the grouping columns are not passed to the applied function, so a function
+    # reading `df.b` would fail on the `groupby(["a", "b"])` case before `parallel_apply`
+    # is even reached.
     def func(df):
         dum = 0
-        for item in df.b:
+        for item in df.c:
             dum += math.log10(math.sqrt(math.exp(item**2)))
 
-        return dum / len(df.b)
+        return dum / len(df.c)
 
     return func