Add groupby::replace_nulls(replace_policy) api

[isVoid]

Part 1 of #4896, follow up of #6907

This PR provides a groupby version of the replace_nulls(replace_policy) function. A regular replace_nulls(replace_policy) operation updates the nulls with the first non-null value that precedes/follows the null. The groupby version is similar, with an exception that the non-null value look-up is bounded by groups.

Here is an example to illustrate the API input/output behavior:

#Input:
keys = [2, 1, 2, 1]
values = [3, 4, NULL, NULL]

#Output, group order is not guaranteed:
sorted_keys = [1, 1, 2, 2]
result = [4, 4, 3, 3]

[isVoid]

Requires NVIDIA/thrust#1374

[jrhemstad]

This approach seems sub-optimal to me as it requires deep-copying the group labels just to do a groupby fillna. If instead you could request the fillna directly from the groupby object you could avoid that deep copy.

[isVoid]

This approach seems sub-optimal to me as it requires deep-copying the group labels just to do a groupby fillna. If instead you could request the fillna directly from the groupby object you could avoid that deep copy.

I thought about both - current impl seems less intrusive to the groupby API. I was looking at groupby::get_groups and assumed retrieving groupby meta information is a common practice. But it certainly has performance concerns. Will address on next commit.

[harrism]

@isVoid can you add more clarity to the PR description? In particular, I don't understand what "segmented null replacement" means. Are you replacing nulls in only some groups? If it is all groups, then I would think a regular "flat" null replacement would work.

[isVoid]

@isVoid can you add more clarity to the PR description? In particular, I don't understand what "segmented null replacement" means. Are you replacing nulls in only some groups? If it is all groups, then I would think a regular "flat" null replacement would work.

@harrism Yes, it includes all groups. Does this look good?

[harrism]

Since the documentation is not written yet, I still don't understand, but I'll wait.

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (branch-21.06@0ebf7e6). Click here to learn what that means.
The diff coverage is n/a.

@@               Coverage Diff               @@
##             branch-21.06    #7118   +/-   ##
===============================================
  Coverage                ?   82.89%           
===============================================
  Files                   ?      105           
  Lines                   ?    17875           
  Branches                ?        0           
===============================================
  Hits                    ?    14818           
  Misses                  ?     3057           
  Partials                ?        0           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0ebf7e6...7a90f52. Read the comment docs.

[nvdbaranec]

Couple of very tiny things.

[brandon-b-miller]

Still scratching my head to make sure but I think this might unnecessarily materialize a RangeIndex.

[isVoid]

It might, but it depends on whether the upstream passes in a RangeIndex in values. Here the meaning of values are the value columns in the groupby.replace_nulls operation, RangeIndex shouldn't exist. (Python interface is yet to be added so it is a bit hard to see).

[brandon-b-miller]

One question otherwise cython LGTM

[jrhemstad]

@gpucibot merge