DOC: add section on bisect GH35685

Author: seljaks

doc/source/development/maintaining.rst

@@ -121,6 +121,46 @@ Here's a typical workflow for triaging a newly opened issue.
    unless it's know that this issue should be addressed in a specific release (say
    because it's a large regression).
 
+.. _maintaining.regressions:
+
+Investigating regressions
+-------------------------
+
+Regressions are bugs that unintentionally break previously working pandas code.
+The common way to investigate regressions is by using `git bisect <https://git-scm.com/docs/git-bisect>`_,
+which finds the first commit after the bug appears.
+
+For example: a user reports that ``pd.Series([1, 1]).sum()`` returns ``3``
+in pandas version ``1.5.0`` while in version ``1.4.0`` it returned ``2``. To begin,
+create a file ``t.py`` in your pandas directory, which contains
+
+.. code-block:: python
+
+    import pandas as pd
+    assert pd.Series([1, 1]).sum() == 2
+
+and then run::
+
+    git bisect start
+    git bisect good v1.4.0
+    git bisect bad v1.5.0
+    git bisect run bash -c "python setup.py build_ext -j 4; python t.py"
+
+This finds the first commit that changed the behavior. To finish, exit bisect and
+rebuilt the latest C extensions with::
+
+    git bisect reset
+    python setup.py build_ext -j 4
+
+Lastly, report your findings under the corresponding issue and ping the commit author to
+get their input.
+
+.. note::
+    The ``run`` command above will treat commits where ``t.py`` exits with ``0`` as good and other
+    exits as bad. When raising an error is the desired behavior, wrap the code in an appropriate
+    ``try/except`` statement. See `GH35650 <https://github.com/pandas-dev/pandas/issues/35685>`_ for
+    more examples.
+
 .. _maintaining.closing:
 
 Closing issues