[SPARK-17842] Thread and memory leak in WindowDstream (UnionRDD ) when parallelPartition computation gets enabled.

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Duplicate
• Affects Version/s: 2.0.0
• Component/s: DStreams, Spark Core

Environment

Yarn cluster, Eclipse Dev Env

• Assignee: Unassigned
• Reporter: Sreelal S L
• Created: 09/Oct/16 22:15
• Updated: 25/Oct/16 06:53
• Resolved: 10/Oct/16 08:43

Description

We noticed a steady increase in ForkJoinTask instances in the driver process heap. Found out the root cause to be UnionRDD.

WindowDstream internally uses UnionRDD which has a parallel partition computation logic by using parallel collection with ForkJoinPool task support.
partitionEvalTaskSupport =new ForkJoinTaskSupport(new ForkJoinPool(8))

The pool is created each time when a UnionRDD is created , but the pool is not getting shutdown. This is leaking thread/mem every slide interval of the window.

Easily reproducible with the below code. Just keep a watch on the number of threads.

    val sparkConf = new SparkConf().setMaster("local[*]").setAppName("TestLeak")
    val ssc = new StreamingContext(sparkConf, Seconds(1))
    ssc.checkpoint("checkpoint")
    val rdd = ssc.sparkContext.parallelize(List(1,2,3))
    val constStream = new ConstantInputDStream[Int](ssc,rdd)
    constStream.window(Seconds(20),Seconds(1)).print()
    ssc.start()
    ssc.awaitTermination();

This happens only when the number of rdds to be unioned is above the value spark.rdd.parallelListingThreshold (By default 10)

Currently i'm working around by setting this threshold be a higher value.

Issue Links

• duplicates: SPARK-17396 Threads number keep increasing when query on external CSV partitioned table