[
https://issues.apache.org/jira/browse/SPARK-13747?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16001365#comment-16001365
]
Shixiong Zhu commented on SPARK-13747:
--------------------------------------
[~mousa] This is is because Spark uses ThreadLocal in a fork-join pool. Let me
try to clarify the issue.
A fork-join pool allows to run another pending task in the same thread when a
running task is calling Await.ready/result. The magic is when someone calls
Await.ready/result, it will first check if there is any pending task submitted
to the pool, if so, it will call the pending task instead of waiting. (See
scala.concurrent#blocking)
In Spark, the codes hitting this issue have the following pattern.
{code}
try {
check if a thread local is set
if so, throw an exception
else
set the thread local value
do some work
Call Await.ready/result to wait for a result // This doesn't clear the
thread local value.
// If the fork-join pool
schedules a pending task here,
// it will see the thread
local value.
do some work
} finally {
clear the thread local value.
}
{code}
My PR is basically just not calling `scala.concurrent#blocking`. It just makes
a fork-join pool become a normal thread pool executor.
> Concurrent execution in SQL doesn't work with Scala ForkJoinPool
> ----------------------------------------------------------------
>
> Key: SPARK-13747
> URL: https://issues.apache.org/jira/browse/SPARK-13747
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 2.0.0, 2.0.1
> Reporter: Shixiong Zhu
> Assignee: Shixiong Zhu
>
> Run the following codes may fail
> {code}
> (1 to 100).par.foreach { _ =>
> println(sc.parallelize(1 to 5).map { i => (i, i) }.toDF("a", "b").count())
> }
> java.lang.IllegalArgumentException: spark.sql.execution.id is already set
> at
> org.apache.spark.sql.execution.SQLExecution$.withNewExecutionId(SQLExecution.scala:87)
>
> at
> org.apache.spark.sql.DataFrame.withNewExecutionId(DataFrame.scala:1904)
> at org.apache.spark.sql.DataFrame.collect(DataFrame.scala:1385)
> {code}
> This is because SparkContext.runJob can be suspended when using a
> ForkJoinPool (e.g.,scala.concurrent.ExecutionContext.Implicits.global) as it
> calls Await.ready (introduced by https://github.com/apache/spark/pull/9264).
> So when SparkContext.runJob is suspended, ForkJoinPool will run another task
> in the same thread, however, the local properties has been polluted.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
