Improve the performance of filter functions

[hackeryang]

Description

In the original implementation of ArrayFilterFunction.filter*, an additional copy of the filtered block is adopted. The speed, especially for some nested column types such as RowType is not perfect.

So we improved the implementation to record the block index position, then directly call Block.copyPositions() after the filtering is completed, in order to reduce unnecessary block copies such as Type.appendTo() and BlockBuilder.build().

Additional context and related issues

Before the change

Benchmark                             (name)  Mode  Cnt   Score   Error  Units  
BenchmarkArrayFilter.benchmark        filter  avgt   20  22.543 ± 0.979  ns/op  
BenchmarkArrayFilter.benchmarkObject  filter  avgt   20  42.045 ± 2.088  ns/op 

After the change

Benchmark                             (name)  Mode  Cnt   Score   Error  Units  
BenchmarkArrayFilter.benchmark        filter  avgt   20  13.576 ± 0.849  ns/op  
BenchmarkArrayFilter.benchmarkObject  filter  avgt   20  32.051 ± 0.268  ns/op 

Release notes

( ) This is not user-visible or docs only and no release notes are required.
(x) Release notes are required, please propose a release note for me.
( ) Release notes are required, with the following suggested text:

# General
*  Improve performance of `filter` function on arrays. ({issue}`16681`)

[hackeryang]

Hello @findepi @ebyhr @sopel39 , can you please take a review when you have time, thank you~

[raunaqmorarka]

Could you check if writing this part as

positions[length] = position;
length += TRUE.equals(keep) ? 1 : 0;

would give better results ?

[hackeryang]

I reverted to the if clause and benchmarked again, the result of if clause was below:

Benchmark                             (name)  Mode  Cnt   Score   Error  Units  
BenchmarkArrayFilter.benchmark        filter  avgt   20  13.320 ± 0.610  ns/op  
BenchmarkArrayFilter.benchmarkObject  filter  avgt   20  32.699 ± 0.945  ns/op 

The result of ? : clause was below:

Benchmark                             (name)  Mode  Cnt   Score   Error  Units  
BenchmarkArrayFilter.benchmark        filter  avgt   20  13.576 ± 0.849  ns/op  
BenchmarkArrayFilter.benchmarkObject  filter  avgt   20  32.051 ± 0.268  ns/op  

It seems that the performance didn't have too much improvement contrast to the if clause, but i still changed to this way because the code is cleaner~

[raunaqmorarka]

I think whether you see improvement depends on whether the input data has a significant fraction of randomly generated nulls

[hackeryang]

I think whether you see improvement depends on whether the input data has a significant fraction of randomly generated nulls

@raunaqmorarka Good question, you reminded me~ So I temporarily changed the nullRate to 0.8F in BlockAssertions#createRandomBlockForType(the former null rate was 20%):

Then i benchmarked again, below was the result(i.e. with 80% null values):

Before our improvement

Benchmark                             (name)  Mode  Cnt   Score   Error  Units
BenchmarkArrayFilter.benchmarkObject  filter  avgt   20   7.694 ± 0.345  ns/op

After our improvement

Benchmark                             (name)  Mode  Cnt   Score   Error  Units
BenchmarkArrayFilter.benchmarkObject  filter  avgt   20   7.301 ± 2.077  ns/op

So our improvement will only have a little speed up, if there are most null values in the column.

But the less null values are, the faster our improvement will be.

[hackeryang]

@raunaqmorarka Thank you for your advice above very much, i have modified some implementations, please review again when you have time~