为什么在调用 collect() 时我的 Spark 工作会卡住？

Question

Am new to spark. I have created a spark job which does some data processing per user. What am trying to do is get all files in a directory and process those files. There are multiple directories and is of multiple users.

After reading files within one users directory I do some transformation post which I need to work on them collectively (like remove some duplicates based on data). For doing this I am calling the collect() over the RDD.

When running this with 10 directories, it works fine, but when running with a 1000 directories it gets stuck at the collect() call.

I have only done local testing.

Initing spark:

private lazy val sparkSession = SparkSession
.builder()
.appName("Custom Job")
.master("local[*]")
.getOrCreate()

Reading directories and parellelizing:

val allDirs: Seq[String] = fs.getAllDirInPath(Configuration.inputDir)
val paths: RDD[String] = SessionWrapper.getSparkContext.parallelize(allDirs)

Transformations and collect call:

paths.foreachPartition { partition =>
      partition.foreach { dir =>
        val dirData = readDataByDir(dir) // RDD[String]
        val transformed = doTranform(dirData) // RDD[CustomObject]
        val collectedData = tranformed.collect()
        // Do something on collected data
        writeToFile(collectedData)
      }
    }

Some logs from the stuck console:

20/09/09 19:24:40 INFO SparkContext: Starting job: collect at MyCustomHelperWithCollectCall.scala:18
20/09/09 19:24:40 INFO FileInputFormat: Total input paths to process : 3
20/09/09 19:24:40 INFO CombineFileInputFormat: DEBUG: Terminated node allocation with : CompletedNodes: 1, size left: 65935
20/09/09 19:24:40 INFO DAGScheduler: Got job 2 (collect at MyCustomHelperWithCollectCall.scala:18) with 2 output partitions
20/09/09 19:24:40 INFO DAGScheduler: Final stage: ResultStage 2 (collect at MyCustomHelperWithCollectCall.scala:18)
20/09/09 19:24:40 INFO DAGScheduler: Parents of final stage: List()
20/09/09 19:24:40 INFO SparkContext: Starting job: collect at MyCustomHelperWithCollectCall.scala:18
20/09/09 19:24:40 INFO SparkContext: Starting job: collect at MyCustomHelperWithCollectCall.scala:18
20/09/09 19:24:40 INFO DAGScheduler: Missing parents: List()
20/09/09 19:24:40 INFO SparkContext: Starting job: collect at MyCustomHelperWithCollectCall.scala:18
20/09/09 19:24:40 INFO SparkContext: Starting job: collect at MyCustomHelperWithCollectCall.scala:18
20/09/09 19:24:40 INFO SparkContext: Starting job: collect at MyCustomHelperWithCollectCall.scala:18
20/09/09 19:24:40 INFO DAGScheduler: Submitting ResultStage 2 (MapPartitionsRDD[102] at map at MyCustomHelperWithCollectCall.scala:18), which has no missing parents
20/09/09 19:24:40 INFO SparkContext: Starting job: collect at MyCustomHelperWithCollectCall.scala:18
20/09/09 19:24:40 INFO SparkContext: Starting job: collect at MyCustomHelperWithCollectCall.scala:18
20/09/09 19:24:40 INFO SparkContext: Starting job: collect at MyCustomHelperWithCollectCall.scala:18
20/09/09 19:24:40 INFO SparkContext: Starting job: collect at MyCustomHelperWithCollectCall.scala:18
20/09/09 19:24:40 INFO MemoryStore: Block broadcast_14 stored as values in memory (estimated size 7.2 KiB, free 2001.0 MiB)
20/09/09 19:24:40 INFO SparkContext: Starting job: collect at MyCustomHelperWithCollectCall.scala:18
20/09/09 19:24:40 INFO MemoryStore: Block broadcast_14_piece0 stored as bytes in memory (estimated size 3.5 KiB, free 2001.0 MiB)
20/09/09 19:24:40 INFO BlockManagerInfo: Added broadcast_14_piece0 in memory on 192.168.31.222:55666 (size: 3.5 KiB, free: 2004.3 MiB)
20/09/09 19:24:40 INFO SparkContext: Created broadcast 14 from broadcast at DAGScheduler.scala:1200
20/09/09 19:24:40 INFO DAGScheduler: Submitting 2 missing tasks from ResultStage 2 (MapPartitionsRDD[102] at map at MyCustomHelperWithCollectCall.scala:18) (first 15 tasks are for partitions Vector(0, 1))
20/09/09 19:24:40 INFO TaskSchedulerImpl: Adding task set 2.0 with 2 tasks
20/09/09 19:24:40 INFO DAGScheduler: Got job 3 (collect at MyCustomHelperWithCollectCall.scala:18) with 1 output partitions
20/09/09 19:24:40 INFO DAGScheduler: Final stage: ResultStage 3 (collect at MyCustomHelperWithCollectCall.scala:18)
20/09/09 19:24:40 INFO DAGScheduler: Parents of final stage: List()
20/09/09 19:24:40 INFO DAGScheduler: Missing parents: List()
20/09/09 19:24:40 INFO DAGScheduler: Submitting ResultStage 3 (MapPartitionsRDD[96] at map at MyCustomHelperWithCollectCall.scala:18), which has no missing parents
20/09/09 19:24:40 INFO MemoryStore: Block broadcast_15 stored as values in memory (estimated size 7.2 KiB, free 2001.0 MiB)
20/09/09 19:24:40 INFO MemoryStore: Block broadcast_15_piece0 stored as bytes in memory (estimated size 3.5 KiB, free 2001.0 MiB)
20/09/09 19:24:40 INFO BlockManagerInfo: Added broadcast_15_piece0 in memory on 192.168.31.222:55666 (size: 3.5 KiB, free: 2004.3 MiB)
20/09/09 19:24:40 INFO SparkContext: Created broadcast 15 from broadcast at DAGScheduler.scala:1200
20/09/09 19:24:40 INFO DAGScheduler: Submitting 1 missing tasks from ResultStage 3 (MapPartitionsRDD[96] at map at MyCustomHelperWithCollectCall.scala:18) (first 15 tasks are for partitions Vector(0))
20/09/09 19:24:40 INFO TaskSchedulerImpl: Adding task set 3.0 with 1 tasks
20/09/09 19:24:40 INFO DAGScheduler: Got job 4 (collect at MyCustomHelperWithCollectCall.scala:18) with 1 output partitions
20/09/09 19:24:40 INFO DAGScheduler: Final stage: ResultStage 4 (collect at MyCustomHelperWithCollectCall.scala:18)
20/09/09 19:24:40 INFO DAGScheduler: Parents of final stage: List()
20/09/09 19:24:40 INFO DAGScheduler: Missing parents: List()
20/09/09 19:24:40 INFO DAGScheduler: Submitting ResultStage 4 (MapPartitionsRDD[101] at map at MyCustomHelperWithCollectCall.scala:18), which has no missing parents
20/09/09 19:24:40 INFO FileInputFormat: Total input paths to process : 5

Please help !

Answer 1

Collect (Action) - Return all the elements of the dataset as an array at the driver program. This is usually useful after a filter or other operation that returns a sufficiently small subset of the data

Seems like you have less memory dont call collect use

df.show(100) 

or

df.take(100)

Also update your Spark DSG graph into question to understand processing