[jira] [Created] (ARROW-9918) [Rust] [DataFusion] Move away from builders to create arrays

["Jorge (Jira)" <ji...@apache.org>]

Jorge created ARROW-9918:
----------------------------

             Summary: [Rust] [DataFusion] Move away from builders to create arrays
                 Key: ARROW-9918
                 URL: https://issues.apache.org/jira/browse/ARROW-9918
             Project: Apache Arrow
          Issue Type: Improvement
          Components: Rust, Rust - DataFusion
            Reporter: Jorge
            Assignee: Jorge


[~yordan-pavlov]  hinted in https://issues.apache.org/jira/browse/ARROW-8908 that there is a performance hit when using array builders.

Indeed, I get about 15% speedup by using {{From}} instead of a builder in DataFusion's math operations, and this number _increases with batch size_, which is detrimental since larger batch sizes leverage SDMI better.

Below, {{sqrt_X_Y}}, X is the log2 of the array's size, Y is log2 of the batch size:

 
{code:java}
sqrt_20_12              time:   [34.422 ms 34.503 ms 34.584 ms]                       
                        change: [-16.333% -16.055% -15.806%] (p = 0.00 < 0.05)
                        Performance has improved.

sqrt_22_12              time:   [150.13 ms 150.79 ms 151.42 ms]                       
                        change: [-16.281% -15.488% -14.779%] (p = 0.00 < 0.05)
                        Performance has improved.

sqrt_22_14              time:   [151.45 ms 151.68 ms 151.90 ms]                       
                        change: [-18.233% -16.919% -15.888%] (p = 0.00 < 0.05)
                        Performance has improved.
{code}
See how there is no difference between {{sqrt_20_12}} and {{sqrt_22_12}} (same batch size), but a measurable difference between {{sqrt_22_12}} and {{sqrt_22_14}} (different batch size).

This issue proposes that we migrate our datafusion and rust operations from builder-base array construction to a non-builder based, whenever possible.

 

Most code can be easily migrated to use {{From}} trait, a migration is along the lines of

 
{code:java}
        let mut builder = Float64Builder::new(array.len());
        for i in 0..array.len() {
            if array.is_null(i) {
                builder.append_null()?;
            } else {
                builder.append_value(array.value(i).$FUNC())?;
            }
        }
        Ok(Arc::new(builder.finish()))
{code}
 to

 
{code:java}
        let result: Float64Array = (0..array.len())
            .map(|i| {
                if array.is_null(i) {
                    None
                } else {
                    Some(array.value(i).$FUNC())
                }
            })
            .collect::<Vec<Option<f64>>>()
            .into();
        Ok(Arc::new(result)){code}
and this is even simpler as we do not need to use an extra API (Builder).

 



--
This message was sent by Atlassian Jira
(v8.3.4#803005)