ARROW-12165: [Rust] inline append functions of builders

[ritchie46]

The append functions in the Builder structs are often used in "hot" code. This PR tags them with #[inline], making it possible to inline the function calls across crate boundaries.

[nevi-me]

This makes sense, there's lots of appending going on here 😵

[Dandandan]

@ritchie46 did you have some perf results on this?
I think it definitely makes sense, especially as the functions are wrapped in (unnecessary) Result now which might make things worse. Sometimes inlining makes things worse performance wise, that's why I think we should have some numbers.

[Dandandan]

This seems a large function to inline? Maybe we need to be conservative here and not mark it as inline

[ritchie46]

Ok, I will remove that one.

[ritchie46]

@ritchie46 did you have some perf results on this?
I think it definitely makes sense, especially as the functions are wrapped in (unnecessary) Result now which might make things worse. Sometimes inlining makes things worse performance wise, that's why I think we should have some numbers.

There are quite some functions, but for met the most important one is the primitive builder.

Running this benchmark:

fn create_primitive_array(n: usize) -> PrimitiveArray<Int64Type> {
    let mut builder = PrimitiveBuilder::new(n);

    for i in 0..n {
        builder.append_value(i as i64).unwrap();
    }

    builder.finish()
}

fn add_benchmark(c: &mut Criterion) {
    c.bench_function("512", |b| b.iter(|| create_primitive_array(512)));
    c.bench_function("4096", |b| b.iter(|| create_primitive_array(4096)));
}

Gave the following improvement:

PrimitiveBuilder

Gnuplot not found, using plotters backend
512                     time:   [848.60 ns 848.70 ns 848.81 ns]                 
                        change: [-54.476% -54.424% -54.383%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 6 outliers among 100 measurements (6.00%)
  2 (2.00%) low mild
  1 (1.00%) high mild
  3 (3.00%) high severe

4096                    time:   [5.3370 us 5.3449 us 5.3539 us]                  
                        change: [-62.787% -62.696% -62.617%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 2 outliers among 100 measurements (2.00%)
  1 (1.00%) high mild
  1 (1.00%) high severe

StringBuilder

For a StringBuilder it shows the following results. Still improvements, but less, because the function call is a smaller part of the work.

Gnuplot not found, using plotters backend
512                     time:   [6.1797 us 6.1825 us 6.1851 us]                 
                        change: [-9.0882% -9.0515% -9.0141%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 1 outliers among 100 measurements (1.00%)
  1 (1.00%) high severe

4096                    time:   [37.948 us 37.957 us 37.965 us]                  
                        change: [-13.363% -13.248% -13.149%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 2 outliers among 100 measurements (2.00%)
  1 (1.00%) high mild
  1 (1.00%) high severe

[ritchie46]

especially as the functions are wrapped in (unnecessary) Result now which might make things worse.

Is there a reason not to remove this? Backwards incompatibility changes are already happened, so maybe we can remove this?

[github-actions]

https://issues.apache.org/jira/browse/ARROW-12165

[Dandandan]

@ritchie46 quite nice micro benchmark results 👍

Is there a reason not to remove this? Backwards incompatibility changes are already happened, so maybe we can remove this?

I don't think there is any reason. I tried to do it some time ago, but it requires a lot of work as it is used in quite some code as you can imagine.

[codecov-io]

Codecov Report

Merging #9860 (1cb93ee) into master (4de0ed7) will decrease coverage by 0.00%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master    #9860      +/-   ##
==========================================
- Coverage   82.61%   82.60%   -0.01%     
==========================================
  Files         254      255       +1     
  Lines       59583    59588       +5     
==========================================
  Hits        49225    49225              
- Misses      10358    10363       +5     

Impacted Files	Coverage Δ
rust/arrow/src/array/builder.rs	85.22% <ø> (ø)
rust/parquet/benches/arrow_writer.rs	0.00% <0.00%> (ø)

---

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 7a0d334...1cb93ee. Read the comment docs.

[alamb]

I don't think there is any reason. I tried to do it some time ago, but it requires a lot of work as it is used in quite some code as you can imagine.

Yeah I agree it would make the API easier to work with if the builders didn't return Result. I think it would be worth the backwards compatibility hit but as @Dandandan says it is a lot of work

[alamb]

Thanks @ritchie46 -- this looks great to me