[jira] [Comment Edited] (BEAM-10100) FileIO writeDynamic with AvroIO.sink not writing all data

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

    [ https://issues.apache.org/jira/browse/BEAM-10100?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17117718#comment-17117718 ] 

Dave Martin edited comment on BEAM-10100 at 5/27/20, 12:20 PM:
---------------------------------------------------------------

Thanks [~iemejia]

Ive just tested with DirectRunner and the result is expected (i.e. a full export is achieved - no dropped records). So this seems to be an issue only for SparkRunner.


was (Author: djtfmartin):
Thanks [~iemejia]

Just tested with DirectRunner and the result is expected (i.e. a full export is achieved - no dropped records). So this seems to be an issue only for SparkRunner.

> FileIO writeDynamic with AvroIO.sink not writing all data
> ---------------------------------------------------------
>
>                 Key: BEAM-10100
>                 URL: https://issues.apache.org/jira/browse/BEAM-10100
>             Project: Beam
>          Issue Type: Bug
>          Components: io-java-avro, io-java-files, runner-spark
>    Affects Versions: 2.17.0, 2.20.0
>         Environment: Mac OSX Catalina, tested with SparkRunner - Spark 2.4.5.
>            Reporter: Dave Martin
>            Priority: P2
>
> FileIO writeDynamic with AvroIO.sink is not writing all data in the following pipeline. The amount of data written varies between runs but it is consistently dropping records. This is with a very small test dataset - 6 records, which should produce 3 directories.
> {code:java}
> Pipeline p = Pipeline.create(options);
> PCollection<KV<String, AvroRecord>> records = p.apply(TextIO.read().from("/tmp/input.csv"))
> .apply(ParDo.of(new StringToDatasetIDAvroRecordFcn()));
> //write out into AVRO in each separate directory
> records.apply("Write avro file per dataset", FileIO.<String, KV<String, AvroRecord>>writeDynamic()
>   .by(KV::getKey)
>   .via(Contextful.fn(KV::getValue), Contextful.fn(x -> AvroIO.sink(AvroRecord.class).withCodec(BASE_CODEC)))
>   .to(options.getTargetPath())
>   .withDestinationCoder(StringUtf8Coder.of())
>   .withNaming(key -> defaultNaming(key + "/export", PipelinesVariables.Pipeline.AVRO_EXTENSION)));
> p.run().waitUntilFinish();
> {code}
> If i replace AvroIO.sink() with TextIO.sink() (and replace the initial mapping function) then the correct number of records are written to the separate directories. This is working consistently.
> e.g.
> {code:java}
> // Initialise pipeline
> Pipeline p = Pipeline.create(options);
> PCollection<KV<String, String>> records = p.apply(TextIO.read().from("/tmp/input.csv")).apply(ParDo.of(new StringToDatasetIDKVFcn()));
> //write out into AVRO in each separate directory
> records.apply("Write CSV file per dataset", FileIO.<String, KV<String, String>>writeDynamic()
>     .by(KV::getKey)
>     .via(Contextful.fn(KV::getValue), TextIO.sink())
>     .to(options.getTargetPath())
>     .withDestinationCoder(StringUtf8Coder.of())
>     .withNaming(datasetID -> defaultNaming(key + "/export", ".csv"));
>  p.run().waitUntilFinish();
> {code}
> cc [~timrobertson100]



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