feat: Optimize for CASE WHEN .. THEN column ELSE null END

[andygrove]

Which issue does this PR close?

Part of #669

Rationale for this change

CaseExpr is an expensive expression and there are some cases where can use specialized expressions instead.

For the usage CASE WHEN predicate THEN column ELSE null END which is seen in numerous TPC-DS queries, we can just evaluate the column and then change the null mask based on the predicate, which is a much faster operation (more than 10x faster than using the standard case expression).

What changes are included in this PR?

How are these changes tested?

I manually tested with this query.

select
    sum(case when (ss_quantity=1) then ss_sales_price else null end) sun_sales,
    sum(case when (ss_quantity=2) then ss_sales_price else null end) mon_sales,
    sum(case when (ss_quantity=3) then ss_sales_price else null end) tue_sales,
    sum(case when (ss_quantity=4) then ss_sales_price else null end) wed_sales,
    sum(case when (ss_quantity=5) then ss_sales_price else null end) thu_sales,
    sum(case when (ss_quantity=6) then ss_sales_price else null end) fri_sales,
    sum(case when (ss_quantity=7) then ss_sales_price else null end) sat_sales
from store_sales;

Here are the timings in seconds for five iterations.

Spark

[2.5691187381744385, 1.3875913619995117, 1.2080309391021729, 1.4287035465240479, 1.2525527477264404]

Comet before this PR

[4.076025724411011, 2.6956324577331543, 2.5712761878967285, 2.5598058700561523, 2.5548157691955566]

Comet with this PR

[3.372154951095581, 1.8503100872039795, 1.7506794929504395, 1.722181797027588, 1.696714162826538]

[andygrove]

I also see a ~5% performance improvement with TPC-DS q43 @ sf=100

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 33.74%. Comparing base (de8c55e) to head (49e53d8).
Report is 1 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff              @@
##               main     #672      +/-   ##
============================================
+ Coverage     33.69%   33.74%   +0.04%     
+ Complexity      840      835       -5     
============================================
  Files           109      109              
  Lines         42527    42529       +2     
  Branches       9343     9343              
============================================
+ Hits          14331    14352      +21     
+ Misses        25245    25211      -34     
- Partials       2951     2966      +15     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[andygrove]

@parthchandra @kazuyukitanimura @huaxingao @viirya This is ready for review now

[parthchandra]

Couldn't we have added this as an optimization in DF itself?

[andygrove]

I filed apache/datafusion#11484 to track up streaming this

[andygrove]

DataFusion PR: apache/datafusion#11534

[andygrove]

Let's just wait to get these optimizations from upstream DataFusion. I can test against the lastest DF code when benchmarking.