[
https://issues.apache.org/jira/browse/SPARK-23124?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Apache Spark reassigned SPARK-23124:
------------------------------------
Assignee: Apache Spark
> Warn users when broacast big table in JoinSelection instead of just run it
> --------------------------------------------------------------------------
>
> Key: SPARK-23124
> URL: https://issues.apache.org/jira/browse/SPARK-23124
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 2.1.0, 2.3.0
> Reporter: liupengcheng
> Assignee: Apache Spark
> Priority: Major
>
> When running a SparkSQL thritserver, we encountered sudden corruption of the
> thritserver which is caused by OutOfMemoryError.
> After review the code and some debug, I finally find out that the framework
> permit broadcast big table and give no warnings, detail code see below:
> {code:java}
> case logical.Join(left, right, joinType, condition) =>
> val buildSide = broadcastSide(canBuildLeft = true, canBuildRight = true,
> left, right)
> // This join could be very slow or OOM
> joins.BroadcastNestedLoopJoinExec(
> planLater(left), planLater(right), buildSide, joinType, condition) :: Nil
> private def broadcastSide(
> canBuildLeft: Boolean,
> canBuildRight: Boolean,
> left: LogicalPlan,
> right: LogicalPlan): BuildSide = {
> def smallerSide =
> if (right.stats.sizeInBytes <= left.stats.sizeInBytes) BuildRight else
> BuildLeft
> val buildRight = canBuildRight && right.stats.hints.broadcast
> val buildLeft = canBuildLeft && left.stats.hints.broadcast
> if (buildRight && buildLeft) {
> // Broadcast smaller side base on its estimated physical size
> // if both sides have broadcast hint
> smallerSide
> } else if (buildRight) {
> BuildRight
> } else if (buildLeft) {
> BuildLeft
> } else if (canBuildRight && canBuildLeft) {
> // for the last default broadcast nested loop join
> smallerSide
> } else {
> throw new AnalysisException("Can not decide which side to broadcast for
> this join")
> }
> }
> {code}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
