In #6996 we ensured that CI passes with pandas 1.1, 1.2 and master version, but I still skipped some tests or ignored warnings, just to get CI working for now.

The open issues to resolve from that were discovered there:

Changes related to pandas 1.1:

• There are still some describe() related warnings ignored in dask/dataframe/tests/test_dataframe.py. The pandas method deprecated behaviour regarding datetimes (seeing them as numeric instead of categorical), and added an option to get the new behaviour (see API: Make describe changes backwards compatible pandas-dev/pandas#34798).
  Question here is whether dask follows that behaviour change of pandas. [ Add datetime_is_numeric to dataframe.describe #7719 ]

Changes related to pandas 1.2.0:

• Pandas 1.2.0 compatibility - std() for the datetime64[ns] dtype #8214
  Some tests are skipped for this in dask/dataframe/tests/test_arithmetics_reduction.py
• Categorical.set_levels(inplace=True) is deprecated in pandas, and is used here in the groupby code.
  Warnings are ignored in dask/dataframe/tests/test_groupby.py [Pandas compat #7712]
• Pandas 1.2 deprecated using a single string as "partial datetime slice" in __getitem__ (eg df["2011-01-02"], but partial strings in an actual slice like df["2011":"2012] still works fine, and using that string with loc as well, i.e. df.loc["2011-01-02"], where there is no ambiguity about selecting a column or slicing the rows).
  The question is if dask wants to follow this change (I assume yes). Currently some tests are skipped in dask/dataframe/tests/test_indexing.py. [WIP Warn if using pandas datetimelike string in dataframe.__getitem__ #7749]
• Index.is_all_datesis deprecated in pandas, and this is used here in the rolling code.
  Warning are ignored in dask/dataframe/tests/test_rolling.py
• Pandas 1.2.0 compatibility - column reductions are applied column-wise (when possible) #8216
  The question here is whether dask wants to follow that behaviour (I am not familiar with how column reductions are currently implemented).

Changes related to pandas master (future 1.3.0):

• Some changes in the rolling code (related to win_type) in pandas will need adjustments in dask to make it compatible, see DEPR: Rolling.win_type returning freq & is_datetimelike pandas-dev/pandas#38963
  Some tests are skipped and warning ignored in dask/dataframe/tests/test_rolling.py [WIP Fix rolling for compatibility with pandas > 1.3.0 #8150]