cudf: [FEA] Behavior of apply_rows on dataframes with missing values

I’m not sure if this is intended behavior or not. Right now it’s a little annoying to use apply_rows on dataframes with missing values. This is the current behavior:

>>> cudf.__version__
'0.8.0+0.g8fa7bd3.dirty'
>>> df = cudf.DataFrame({'a': [1., 2., None, 3.]})
>>> print(df)
     a
0  1.0
1  2.0
2
3  3.0
>>> def kernel(a, out):
...     for i, val in enumerate(a):
...         out[i] = val
>>> df = df.apply_rows(kernel, ['a'], {'out': float}, {})
>>> print(df)
     a       out
0  1.0       1.0
1  2.0       2.0
2            3.0
3  3.0  1.5e-323

You can see that the loop index in the kernel skipped the rows of a with missing values, but the same index applied to out did not skip those rows. So the out rows are out of order. To get what I really want, I have to first mask out the missing values from a, apply the kernel, then apply it back to the original dataframe with some sort of merge. That is, I would want

     a  out
0  1.0  1.0
1  2.0  2.0
2            
3  3.0  3.0

It would be nice if

1. The automatic masking out of missing values in the kernel also applied to the output columns.
2. The output column was initialized with missing values, rather than whatever happened to be in memory, or at least missing values were applied to out for the indices where a has missing values.

I don’t know how missing values are implemented in cudf so it’s possible 2 might be considered too inefficient to do by default.