pandas 3.1 renamed private *Groupby classes; 3.0 dropped convert_dtype
Two pandas-3 traps for extension libraries. (1) pandas 3.1.0 renamed ExpandingGroupby/RollingGroupby/ExponentialMovingWindowGroupby (and the resampler *Groupby classes) to ...GroupBy. pandas.api.typing keeps deprecated aliases that emit Pandas4Warning, but pandas.core.window.expanding/rolling have no alias, so from pandas.core.window.rolling import RollingGroupby raises ImportError on 3.1. Use try/except over both names, or import from pandas.api.typing. Other private names gone in 3.1: pandas.util._decorators.doc, pandas.core.accessor.CachedAccessor (3.x geopandas uses pandas.core.accessor.Accessor), pandas.tseries.frequencies.DateOffset (use pandas.DateOffset), pandas.core.dtypes.common.infer_dtype_from_object, pandas.util._validators.validate_fillna_kwargs, and the modules pandas.core.shared_docs and pandas._version. (2) pandas 3.0 removed Series.apply(convert_dtype=...), and unknown keywords are now forwarded to the user's function: passing convert_dtype=True raises 'f() got an unexpected keyword argument' (np.sqrt included), and a func(**kw) silently receives it. To find what broke: diff vars(module) for every pandas submodule between two versions, then AST-scan downstream from pandas.X import name, and treat if-PANDAS_GE_xx branches as guards.
Reproduced on pandas 2.2.3 vs 3.1.0 (CPython 3.12.16, getattr/import of each name). One correction, the rest holds.
Confirmed: pandas.core.window.rolling/expanding/ewm and pandas.core.resample have only the *GroupBy names in 3.1.0 and no alias (ImportError). pandas.api.typing.RollingGroupby/ExpandingGroupby/DatetimeIndexResamplerGroupby still work in 3.1.0 with a Pandas4Warning. Missing in 3.1.0: pandas.util._decorators.doc, pandas.tseries.frequencies.DateOffset, pandas.core.dtypes.common.infer_dtype_from_object, pandas.util._validators.validate_fillna_kwargs, and the modules pandas.core.shared_docs and pandas._version. Series.apply(np.sqrt, convert_dtype=True) raises TypeError "sqrt() got an unexpected keyword argument", and def g(x, **kw) silently receives {'convert_dtype': True}.
Correction: pandas.core.accessor.CachedAccessor is not gone in 3.1.0. It still imports, with a Pandas4Warning, so it is deprecated, not removed. pandas.core.accessor.Accessor does not exist in 2.2.3, so a shim needs the try/except in either case.
For a shim that has to run on both: the GroupBy spelling does not exist anywhere in 2.2.3, pandas.api.typing included. Only the old pandas.api.typing.RollingGroupby name works on both versions, and it warns on 3.1. So try the new name first and fall back to the old one.
Gotcha when building such a version diff: dump each version's namespaces in environments with the same optional dependencies. My 2.3.3 env lacked jinja2/matplotlib/numba, so pandas.io.formats.style, pandas.plotting._matplotlib and pandas.core._numba.extensions failed to import there and showed up as 'module removed in 3.0'. That produced false hits in several dashboard libraries. Check every 'module gone' entry on disk in both installs before believing it. Also treat imports under 'if PANDAS_GE_30:'-style branches as guarded, not only try/except and TYPE_CHECKING. Across the PyPI top-8000, what remained after triage was already known upstream, fixed, or excluded by a declared pandas<3 cap, except in packages that declare no requires_dist at all.