Column assignments fails with Float64Dtype type

[nils-braun]

What happened:
When using pandas' new Float64 nullable type (with pandas >= 1.2), column assignment fails with

TypeError: Cannot interpret 'Float64Dtype()' as a data type

Minimal Complete Verifiable Example:

Fails at least with pandas version 1.2.0 (below that, the new extension type was not introduced).

import pandas as pd
import dask.dataframe as dd

# some example data. Important is only the Float64, the new pandas extension type
df = dd.from_pandas(pd.DataFrame({"a": [1.1]}, dtype="Float64"), npartitions=1)

df.assign(new_col=df["a"])
# TypeError: Cannot interpret 'Float64Dtype()' as a data type

Full stacktrace

Traceback (most recent call last):
  File "bla.py", line 7, in <module>
    df.assign(new_col=df["a"])
  File "/home/nils/anaconda3/envs/dask-sql/lib/python3.8/site-packages/dask/dataframe/core.py", line 4004, in assign
    df2 = self._meta_nonempty.assign(**_extract_meta(kwargs, nonempty=True))
  File "/home/nils/anaconda3/envs/dask-sql/lib/python3.8/site-packages/dask/dataframe/core.py", line 360, in _meta_nonempty
    return meta_nonempty(self._meta)
  File "/home/nils/anaconda3/envs/dask-sql/lib/python3.8/site-packages/dask/utils.py", line 509, in __call__
    return meth(arg, *args, **kwargs)
  File "/home/nils/anaconda3/envs/dask-sql/lib/python3.8/site-packages/dask/dataframe/utils.py", line 391, in meta_nonempty_dataframe
    dt_s_dict[dt] = _nonempty_series(x.iloc[:, i], idx=idx)
  File "/home/nils/anaconda3/envs/dask-sql/lib/python3.8/site-packages/dask/dataframe/utils.py", line 578, in _nonempty_series
    data = np.array([entry, entry], dtype=dtype)
TypeError: Cannot interpret 'Float64Dtype()' as a data type

As a fix, I guess one wants to introduce a function similar to https://github.com/dask/dask/blob/master/dask/dataframe/utils.py#L43 (but for floats) and use it in https://github.com/dask/dask/blob/master/dask/dataframe/utils.py#L540. I am happy to open a PR to fix this if this solution is ok.

Environment:

• Dask version: 2021.1.1
• Pandas version: 1.2.0
• Python version:
• Operating System: ubuntu
• Install method (conda, pip, source): conda

[jsignell]

I think the solution that you propose is very reasonable and if you are willing to open a PR even better! Just as a heads up there is a release planned for friday (tomorrow) so it'd be great to get this in.

[nils-braun]

@jsignell PR is open - I hope it was still on time yesterday evening (European time)