[BUG]: Rolling.sum() calculated wrong values when axis is one and dtypes are mixed

[phofl]

• closes pandas.DataFrame.rolling produces incorrect value if integer and float objects are mixed #20649
• closes How to stop a rolling window at nan values and continue after it?  #35596
• tests added / passed
• passes black pandas
• passes git diff upstream/master -u -- "*.py" | flake8 --diff
• whatsnew entry

In case of axis=1 and mixed dtypes, the _apply_blockwise did not calculate the sum for a complete row. Through converting the obj and consolidating the blocks we can avoid this.

[mroeschke]

Does this also potentially close #35596?

[phofl]

Thanks, I missed that issue. When the output you posted is the correct one, then yes, this fixes the issue. I will add a test.

[mroeschke]

Could you add a comment with the issue number and a small explanation on why we are doing this?

[phofl]

Done

[jreback]

cc @jbrockmendel FYI

[jreback]

can you add another test which has mixed types with say a string in one column (and also try with datetimes we don't want to coerce), so this raises I think we are going to have an odd error message and may want to catch / re-raise if that is the case

actually i think we need to explicity get outnumeric dtypes (e.g. do a select_dtypes(include=[['integer', 'float']])

otherwise we might include timedelta for example (and we don't know how to coerce back I think)

[jreback]

cc @jbrockmendel FYI

[phofl]

@jreback so we should raise if we get dtypes except float and int?

[mroeschke]

I think we should just drop columns that are not coercible to float like axis=0 behavior

In [1]: df = pd.DataFrame({'a': range(5), 'b': ['a'] * 5})

In [2]: df
Out[2]:
   a  b
0  0  a
1  1  a
2  2  a
3  3  a
4  4  a

In [3]: df.rolling(2).sum()
Out[3]:
     a
0  NaN
1  1.0
2  3.0
3  5.0
4  7.0

[phofl]

Thanks @mroeschke. Implemented it.

[jreback]

lgtm ping on green.

[phofl]

@jreback green

[jreback]

thanks @phofl very nice!