[SPARK-17381] Memory leak org.apache.spark.sql.execution.ui.SQLTaskMetrics

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Not A Problem
• Affects Version/s: 2.0.0
• Component/s: SQL

Environment

EMR 5.0.0 (submitted as yarn-client)
Java Version 1.8.0_101 (Oracle Corporation)
Scala Version version 2.11.8

Problem also happens when I run locally with similar versions of java/scala. OS: Ubuntu 16.04

• Assignee: Unassigned
• Reporter: Joao Duarte
• Created: 02/Sep/16 15:24
• Updated: 19/Jun/17 05:33
• Resolved: 20/Sep/16 14:17

Description

I am running a Spark Streaming application from a Kinesis stream. After some hours running it gets out of memory. After a driver heap dump I found two problems:
1) huge amount of org.apache.spark.sql.execution.ui.SQLTaskMetrics (It seems this was a problem before:
https://issues.apache.org/jira/browse/SPARK-11192);

To replicate the org.apache.spark.sql.execution.ui.SQLTaskMetrics leak just needed to run the code below:

    val dstream = ssc.union(kinesisStreams)
    dstream.foreachRDD((streamInfo: RDD[Array[Byte]]) => {
      val toyDF = streamInfo.map(_ =>
        (1, "data","more data "
        ))
        .toDF("Num", "Data", "MoreData" )
      toyDF.agg(sum("Num")).first().get(0)
    }
    )

2) huge amount of Array[Byte] (9Gb+)

After some analysis, I noticed that most of the Array[Byte] where being referenced by objects that were being referenced by SQLTaskMetrics. The strangest thing is that those Array[Byte] were basically text that were loaded in the executors, so they should never be in the driver at all!

Still could not replicate the 2nd problem with a simple code (the original was complex with data coming from S3, DynamoDB and other databases). However, when I debug the application I can see that in Executor.scala, during reportHeartBeat(), the data that should not be sent to the driver is being added to "accumUpdates" which, as I understand, will be sent to the driver for reporting.

To be more precise, one of the taskRunner in the loop "for (taskRunner <- runningTasks.values().asScala)" contains a GenericInternalRow with a lot of data that should not go to the driver. The path would be in my case: taskRunner.task.metrics.externalAccums[2]._list[0]. This data is similar (if not the same) to the data I see when I do a driver heap dump.

I guess that if the org.apache.spark.sql.execution.ui.SQLTaskMetrics leak is fixed I would have less of this undesirable data in the driver and I could run my streaming app for a long period of time, but I think there will always be some performance lost.