[jira] [Commented] (SPARK-20050) Kafka 0.10 DirectStream doesn't commit last processed batch's offset when graceful shutdown

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

    [ https://issues.apache.org/jira/browse/SPARK-20050?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17500321#comment-17500321 ] 

James Grinter commented on SPARK-20050:
---------------------------------------

You're just working around a genuine bug, and behaviour that is documented as how you're supposed to do it (and using the offsets is a better way, as monitoring tools can then see the consumer-group progress and detect that there's a problem, instead of it being hidden away in a private implementation).

> Kafka 0.10 DirectStream doesn't commit last processed batch's offset when graceful shutdown
> -------------------------------------------------------------------------------------------
>
>                 Key: SPARK-20050
>                 URL: https://issues.apache.org/jira/browse/SPARK-20050
>             Project: Spark
>          Issue Type: Bug
>          Components: DStreams
>    Affects Versions: 2.2.0
>            Reporter: Sasaki Toru
>            Priority: Major
>
> I use Kafka 0.10 DirectStream with properties 'enable.auto.commit=false' and call 'DirectKafkaInputDStream#commitAsync' finally in each batches,  such below
> {code}
> val kafkaStream = KafkaUtils.createDirectStream[String, String](...)
> kafkaStream.map { input =>
>   "key: " + input.key.toString + " value: " + input.value.toString + " offset: " + input.offset.toString
>   }.foreachRDD { rdd =>
>     rdd.foreach { input =>
>     println(input)
>   }
> }
> kafkaStream.foreachRDD { rdd =>
>   val offsetRanges = rdd.asInstanceOf[HasOffsetRanges].offsetRanges
>   kafkaStream.asInstanceOf[CanCommitOffsets].commitAsync(offsetRanges)
> }
> {\code}
> Some records which processed in the last batch before Streaming graceful shutdown reprocess in the first batch after Spark Streaming restart, such below
> * output first run of this application
> {code}
> key: null value: 1 offset: 101452472
> key: null value: 2 offset: 101452473
> key: null value: 3 offset: 101452474
> key: null value: 4 offset: 101452475
> key: null value: 5 offset: 101452476
> key: null value: 6 offset: 101452477
> key: null value: 7 offset: 101452478
> key: null value: 8 offset: 101452479
> key: null value: 9 offset: 101452480  // this is a last record before shutdown Spark Streaming gracefully
> {\code}
> * output re-run of this application
> {code}
> key: null value: 7 offset: 101452478   // duplication
> key: null value: 8 offset: 101452479   // duplication
> key: null value: 9 offset: 101452480   // duplication
> key: null value: 10 offset: 101452481
> {\code}
> It may cause offsets specified in commitAsync will commit in the head of next batch.



--
This message was sent by Atlassian Jira
(v8.20.1#820001)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org