[jira] [Created] (BEAM-6064) Python BigQuery performance much worse than Java

["Jan Kuipers (JIRA)" <ji...@apache.org>]

Jan Kuipers created BEAM-6064:
---------------------------------

             Summary: Python BigQuery performance much worse than Java
                 Key: BEAM-6064
                 URL: https://issues.apache.org/jira/browse/BEAM-6064
             Project: Beam
          Issue Type: Bug
          Components: sdk-py-core
    Affects Versions: 2.8.0
            Reporter: Jan Kuipers
            Assignee: Ahmet Altay


The performance of reading from BigQuery in Python seems to be much worse than the performance of it in Java.

To reproduce this, I've run the following two programs on the Google Cloud, which basically read the weights from the public data set "natality" and outputs the top 100 largest weights.

Python:
{code:java}
# <cut imports>

options = PipelineOptions()
options.view_as(StandardOptions).runner = 'DataflowRunner'
# <cut more options>

pipeline = Pipeline(options=options)
(pipeline
    | 'Read' >> beam.io.Read(beam.io.BigQuerySource(query='SELECT weight_pounds FROM [bigquery-public-data:samples.natality]'))
    | 'MapToFloat' >> beam.Map(lambda elem: elem['weight_pounds'])
    | 'Top' >> beam.combiners.Top.Largest(100)
    | 'MapToString' >> beam.Map(lambda elem: str(elem))
    | 'Write' >> beam.io.WriteToText("<output-file>"))

pipeline.run()
{code}
 Java:
{code:java}
// <cut imports>

public class Natality {
    public static void main(String[] args) {
        DataflowPipelineOptions options = PipelineOptionsFactory.create().as(DataflowPipelineOptions.class);
        options.setRunner(DataflowRunner.class);
        // <cut more options>
        
        Pipeline pipeline = Pipeline.create(options);

        pipeline.apply("Read", BigQueryIO.readTableRows()
            .fromQuery("SELECT weight_pounds FROM [bigquery-public-data:samples.natality]"))
            .apply("MapToDouble", MapElements
                .into(TypeDescriptors.doubles())
                .via(row -> {
                     Object obj = row.get("weight_pounds");
                     return (obj == null ? 0.0 : (Double) obj);
                }))
            .apply("Top", Top.largest(100))
            .apply("MapToString", MapElements
                .into(TypeDescriptors.strings())
                .via(weight -> weight.toString()))
            .apply("Write", TextIO.write().to("<output-file>"));

        pipeline.run().waitUntilFinish();
    }
}
{code}
The "<cut more options>" are basic options like project, job name, temp location, etc. Both programs produce identical outputs.

Running these programs launches a DataFlow job on the Google Cloud with the following results (data from the Google Cloud Platform web interface; I could add screenshots if that were allowed).

Python:
{noformat}
Read Succeeded 1 hr 40 min 40 sec
MapToFloat Succeeded 2 min 43 sec
Top Succeeded 5 min 25 sec
MapToString Succeeded 0 sec
Write Succeeded 3 sec{noformat}
Java:
{noformat}
Read Succeeded 4 min 45 sec
MapToDouble Succeeded 45 sec
Top Succeeded 52 sec
MapToString Succeeded 0 sec
Write Succeeded 1 sec
{noformat}
As you can see, there is an enormous performance hit in Python w.r.t. the reading from BigQuery: 1h40m vs less than 5 minutes.

Furthermore the other standard operations (like Top) are also much slower in Python than in Java.

 



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)