There are already some issues / PRs related to this (eg some upstream failures are also listed in ), but I thought it might be useful to have a dedicated issue about this.

Pandas released a 1.4.0rc0 last week, and probably will release a final 1.4.0 next week.

The upstream test build (eg https://github.com/dask/dask/runs/4823351165?check_suite_focus=true) shows several failures, falling into a few categories:

• Pandas will require a minimum version of sqlalchemy 1.4, which gives import errors in the sql test setup (ImportError: Pandas requires version '1.4.0' or newer of 'sqlalchemy' (version '1.3.23' currently installed).) Moving to SQLAlchemy >= 1.4 #8158
• DataFrame.append is deprecated, and so the warnings are turned into errors. Deprecate append when pandas >= 1.4.0 #8617
• Some timezone related failure with fastparquet (eg test_roundtrip[fastparquet-df12-write_kwargs12-read_kwargs12] is failing)
• test_parquet.py::test_getitem_optimization failures (TypeError: Column name must be a string. Got column None of type NoneType) -> this might also be related to the to_frame failure below Fix to_frame name to not pass None #8554
• dask/dataframe/tests/test_indexing.py::test_to_frame (the result has a "NaN" column name instead of expected default of "0") Fix to_frame name to not pass None #8554
• A set of groupby failures with cumulative methods (test_cumulative) Fix to_frame name to not pass None #8554
• Some errors related to FutureWarning: In a future version, DataFrame.mean(axis=None) will return a scalar mean over the entire DataFrame. To retain the old behavior, use 'frame.mean(axis=0)' or just 'frame.mean()' (test_groupby_column_and_index_apply) Fix axis=None warning #8555
• ufunc related failures (ValueError: Metadata inference failed in __array_wrap__.) caused by pandas removing __array_wrap__ -> this might already be covered by Return da.Array rather than dd.Series for non-ufunc elementwise functions on dd.Series #8558 ?