REF: move towards making _apply_blockwise actually block-wise

[jbrockmendel]

A step towards #34714, orthogonal to the other outstanding PR in this vein #35696.

[jbrockmendel]

cc @simonjayhawkins shouldnt mypy recognize Series as satisfying FrameOrSeries?

pandas/core/window/rolling.py:414: error: Incompatible return value type (got "Series", expected "FrameOrSeries")  [return-value]

[simonjayhawkins]

cc @simonjayhawkins shouldnt mypy recognize Series as satisfying FrameOrSeries?

FrameOrSeries is a typevar, so problems like this involve binding.

so the first error pandas\core\window\rolling.py:414: error: Incompatible return value type (got "Series", expected "FrameOrSeries") [return-value]

FrameOrSeries is the return type of _Window._wrap_results but FrameOrSeries is not bound.

can use FrameOrSeriesUnion as the return type, but if the return type should be the same as passed argument obj then should be bound to obj

i.e.

    def _wrap_results(
        self, results, blocks, obj: FrameOrSeries, exclude=None
    ) -> FrameOrSeries:

but this will still give

pandas\core\window\rolling.py:416: error: Incompatible return value type (got "Series", expected "FrameOrSeries")  [return-value]
pandas\core\window\rolling.py:421: error: Unexpected keyword argument "index" for "NDFrame"  [call-arg]
pandas\core\generic.py:198: note: "NDFrame" defined here
pandas\core\window\rolling.py:421: error: Unexpected keyword argument "columns" for "NDFrame"  [call-arg]
pandas\core\generic.py:198: note: "NDFrame" defined here

since mypy doesn't know that if obj.ndim == 1 refines the type to be Series and other issues need to be addressed since FrameOrSeries is bounded by NDFrame.

for _apply_blockwise, if the return type should be of the same type as self._selected_obj, then should (eventually) create another typevar for this (as FrameOrSeries variable is used for self.obj). However, to bind to a instance variable, need to use Generic, and others won't approve this.

import pandas as pd
from pandas._typing import FrameOrSeries
from typing import Generic


class MyClass(Generic[FrameOrSeries]):
    def __init__(self, obj: FrameOrSeries):
        self._selected_obj = obj

    def foo(self) -> FrameOrSeries:
        return self._selected_obj

reveal_type(MyClass(pd.Series()).foo())
reveal_type(MyClass(pd.DataFrame()).foo())

class MySeries(pd.Series):
    pass

reveal_type(MyClass(MySeries()).foo())

$ mypy ~/mypy-test.py 
C:\Users\simon\mypy-test.py:14: note: Revealed type is 'pandas.core.series.Series*'
C:\Users\simon\mypy-test.py:15: note: Revealed type is 'pandas.core.frame.DataFrame*'
C:\Users\simon\mypy-test.py:20: note: Revealed type is 'mypy-test.MySeries*'

so simplest to get mypy to green is to use FrameOrSeriesUnion for now until we need to refine the types for consistency elsewhere.

[jbrockmendel]

Updated+green. Once this and #35740 are in, Ive got a branch ready that goes all the way and makes _apply_blockwise actually blockwise.

[jreback]

if you can rebsae

[jbrockmendel]

rebased+green