[C++] Add fill_null kernel implementation for (array[string], scalar[string])

[asfimport]

Reporter: Uwe Korn / @xhochy
Assignee: Maarten Breddels / @maartenbreddels

Related issues:

• [C++/Python] Support necessary functionality to have an Arrow-string type in pandas (is depended upon by)

PRs and other links:

• GitHub Pull Request #8628

_{Note: This issue was originally created as ARROW-9489. Please see the migration documentation for further details.}

[asfimport]

Joris Van den Bossche / @jorisvandenbossche:
cc @maartenbreddels depending on time, this would also be an interesting one

In [1]: arr = pa.array(['a', 'b', None])

In [2]: import pyarrow.compute as pc

In [3]: pc.fill_null(arr, 'c')
...
ArrowNotImplementedError: Function fill_null has no kernel matching input types (array[string], scalar[string])

[asfimport]

Maarten Breddels / @maartenbreddels:
Yes, happy to take this on, since it's an ugly code path in Vaex right now, and quite a performance hit.

[asfimport]

Joris Van den Bossche / @jorisvandenbossche:
For full coverage of pandas functionality, a (array[string], array[string]) signature would also be interesting: pandas allows to fill with an array, taking the corresponding value of the second array where the first array has a null.
(but certainly less of a priority)

[asfimport]

Maarten Breddels / @maartenbreddels:
Yes, I thought about that too. Although I think a np.where equivalent would be more general. Sth like:

pc.where(pc.is_null(arr), arr2, arr)  # equivalent what you described
pc.where(pc.is_null(arr), 'c', arr)  # less performant version of fill_null

I could explore what you describe, and see if we can do some template programming so that we can generalize this to other kernels as well. But I think a 'where' kernel would make more sense at this point in time.

[asfimport]

Joris Van den Bossche / @jorisvandenbossche:
Ah, yes, a general "where" general seems more useful indeed, and can be used for this purpose as well (and could also be used for some "setitem" like calls in pandas, which would be useful to have)

[asfimport]

Antoine Pitrou / @pitrou:
Issue resolved by pull request 8628
#8628