How nulls are handled in the boolean mask (selection vector) in a filter kernel varies between languages / data analytics systems (e.g. base R propagates nulls, dplyr R skips (sees as False), SQL generally skips them as well I think, Julia raises an error).

Currently, in Arrow C++ we "propagate" nulls (null in the selection vector gives a null in the output):

In [7]: arr = pa.array([1, 2, 3]) 

In [8]: mask = pa.array([True, False, None]) 

In [9]: arr.filter(mask) 
Out[9]: 
<pyarrow.lib.Int64Array object at 0x7fefe44b3048>
[
  1,
  null
]

Given the different ways this could be done (propagate, skip, error), should we provide an option to control this behaviour?

Reporter: Joris Van den Bossche / @jorisvandenbossche

Related issues:

• [C++] C++ array kernels framework and execution buildout (umbrella issue) (is a child of)

_{Note: This issue was originally created as ARROW-6923. Please see the migration documentation for further details.}