BUG: Series.rolling min_period is ignored and NA behaves strangely

[ghost]

Using current d39a6de

This may be two separate issues, I can't tell.

Case 1.

import pandas as pd
ser=pd.Series(range(10))
ser[:5]=None
print(ser)
pd.Series(ser).rolling(3, min_periods=3).count()

returns

0    0.0
1    0.0
2    0.0
3    0.0
4    0.0
5    1.0
6    2.0
7    3.0
8    3.0
9    3.0
dtype: float64

I expected the first few results to be NA. perhaps related:

pd.Series(ser).rolling(3, min_periods=999).count()

doesn't raise an exception.

Case 2.

for series with DatetimeIndex and rolling with an offset, the behavior is also strange, but different.

import pandas as pd
import random
from itertools import chain
dates = ["2001-01-01"]*2 + ["2001-01-02"]*2 + ["2001-01-03"]*2 + ["2001-01-04"]*2 + ["2001-01-05"]*2+ ["2001-01-06"]*2
ser=pd.Series(index=pd.DatetimeIndex(dates))
ser[0]=111
ser[2]=222
ser[6]=333
res=ser.rolling("2D", min_periods=1).count()
df=pd.DataFrame(dict(data=ser,count=res))
print(df)

result

             data  count
2001-01-01  111.0    1.0   # OK. The day has one non-NA value, 
2001-01-01    NaN    1.0  # and min_period is satisfied
2001-01-02  222.0    2.0  # OK
2001-01-02    NaN    2.0  # OK
2001-01-03    NaN    1.0  # why is this 1.0? this date has no values in it.
2001-01-03    NaN    1.0 # again
2001-01-04  333.0    1.0  # OK
2001-01-04    NaN    1.0 # OK
2001-01-05    NaN    1.0 # why is this 1.0? this date has no values in it.
2001-01-05    NaN    1.0 # again
2001-01-06    NaN    NaN # why is this NaN? min_periods is satisfied
2001-01-06    NaN    NaN # and non-NA count should 0.

encountered in the course of #26959

[WillAyd]

Generally looks like count may be returning 0 instead of NA for non-frequency data. Should return NA as you mention in these cases so would certainly take a PR if you see a way to make it work