[jira] [Commented] (SPARK-6847) Stack overflow on updateStateByKey which followed by a dstream with checkpoint set

["Sean Owen (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-6847?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14489413#comment-14489413 ] 

Sean Owen commented on SPARK-6847:
----------------------------------

I suspect that's exactly it: {{newlist.headOption.orElse(oldstate)}}

Your state has an object graph that grows without bound. Eventually it will cause a problem when, for example, you serialize the state for a checkpoint and the object graph creates an equally deeply nested method call.

I think you simply should not write your state this way, as a lazy evaluation that has to hold on to old state forever. Return either new or old state immediately in the method. I don't think this is a Spark problem.

> Stack overflow on updateStateByKey which followed by a dstream with checkpoint set
> ----------------------------------------------------------------------------------
>
>                 Key: SPARK-6847
>                 URL: https://issues.apache.org/jira/browse/SPARK-6847
>             Project: Spark
>          Issue Type: Bug
>          Components: Streaming
>    Affects Versions: 1.3.0
>            Reporter: Jack Hu
>              Labels: StackOverflowError, Streaming
>
> The issue happens with the following sample code: uses {{updateStateByKey}} followed by a {{map}} with checkpoint interval 10 seconds
> {code}
>     val sparkConf = new SparkConf().setAppName("test")
>     val streamingContext = new StreamingContext(sparkConf, Seconds(10))
>     streamingContext.checkpoint("""checkpoint""")
>     val source = streamingContext.socketTextStream("localhost", 9999)
>     val updatedResult = source.map(
>         (1,_)).updateStateByKey(
>             (newlist : Seq[String], oldstate : Option[String]) =>     newlist.headOption.orElse(oldstate))
>     updatedResult.map(_._2)
>     .checkpoint(Seconds(10))
>     .foreachRDD((rdd, t) => {
>       println("Deep: " + rdd.toDebugString.split("\n").length)
>       println(t.toString() + ": " + rdd.collect.length)
>     })
>     streamingContext.start()
>     streamingContext.awaitTermination()
> {code}
> From the output, we can see that the dependency will be increasing time over time, the {{updateStateByKey}} never get check-pointed,  and finally, the stack overflow will happen. 
> Note:
> * The rdd in {{updatedResult.map(_._2)}} get check-pointed in this case, but not the {{updateStateByKey}} 
> * If remove the {{checkpoint(Seconds(10))}} from the map result ( {{updatedResult.map(_._2)}} ), the stack overflow will not happen



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

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