Spark2.2（三十三）：Spark Streaming和Spark Structured Streaming更新broadcast总结

背景：

需要在spark2.2.0更新broadcast中的内容，网上也搜索了不少文章，都在讲解spark streaming中如何更新，但没有spark structured streaming更新broadcast的用法，于是就这几天进行了反复测试。经过了一下两个测试：：Spark Streaming更新broadcast、Spark Structured Streaming更新broadcast。

1）Spark Streaming更新broadcast（可行）

  def sparkStreaming(): Unit = {
    // Create a local StreamingContext with two working thread and batch interval of 1 second.
    // The master requires 2 cores to prevent a starvation scenario.
    val conf = new SparkConf().setMaster("local[*]").setAppName("NetworkWordCount")
    val ssc = new StreamingContext(conf, Seconds(15))

    // Create a DStream that will connect to hostname:port, like localhost:9999
    val lines = ssc.socketTextStream(ipAddr, 19999)
    val mro = lines.map(row => {
      val fields = row.split(",")
      Mro(fields(0), fields(1))
    })

    val cellJoinMro = mro.transform(row => {
      if (1 < 3) {
        println("更新broadcast..." + new SimpleDateFormat("yyyy-MM-dd HH:mm:ss").format(new java.util.Date()))
        BroadcastWrapper.update(ssc.sparkContext)
      }
      var broadcastCellRes = BroadcastWrapper.getInstance(ssc.sparkContext)
      row.map(row => {
        val int_id: String = row.int_id
        val rsrp: String = row.rsrp
        val findResult: String = String.join(",", broadcastCellRes.value.get(int_id).get)
        val timeStamps: String = String.join(",", findResult)

        CellJoinMro(int_id, rsrp, timeStamps)
      })
    })

    cellJoinMro.print()

    ssc.start() // Start the computation
    ssc.awaitTermination() // Wait for the computation to terminate
  }
import org.apache.spark.SparkContext
import org.apache.spark.broadcast.Broadcast

object BroadcastWrapper {
  @volatile private var instance: Broadcast[Map[String, java.util.List[String]]] = null
  private val baseDir = "/user/my/streaming/test/"

  def loadData(): Map[String, java.util.List[String]] = {
    val files = HdfsUtil.getFiles(baseDir)

    var latest: String = null
    for (key <- files.keySet) {
      if (latest == null) latest = key
      else if (latest.compareTo(key) <= 0) latest = key
    }

    val filePath = baseDir + latest

    val map = HdfsUtil.getFileContent(filePath)
    map
  }

  def update(sc: SparkContext, blocking: Boolean = false): Unit = {
    if (instance != null)
      instance.unpersist(blocking)
    instance = sc.broadcast(loadData())
  }

  def getInstance(sc: SparkContext): Broadcast[Map[String, java.util.List[String]]] = {
    if (instance == null) {
      synchronized {
        if (instance == null) {
          instance = sc.broadcast(loadData)
        }
      }
    }
    instance
  }

}

import java.io.{BufferedReader, InputStreamReader}
import java.text.SimpleDateFormat
import org.apache.hadoop.conf.Configuration
import org.apache.hadoop.fs.Path
import org.apache.hadoop.fs.FileSystem
import scala.collection.mutable

object HdfsUtil {
  private val sdf = new SimpleDateFormat("yyyy-MM-dd 00:00:00")

  def getFiles(path: String): mutable.Map[String, String] = {
    val fileItems = new mutable.LinkedHashMap[String, String]
    val fs = FileSystem.get(new Configuration())
    val files = fs.listStatus(new Path(path))
    var pathStr: String = ""
    for (file <- files) {
      if (file.isFile) {
        pathStr = file.getPath().getName()
        fileItems.put(pathStr.split("/")(pathStr.split("/").length - 1), pathStr)
      }
    }

    fs.close()

    fileItems
  }

  def getFileContent(filePath: String): Map[String, java.util.List[String]] = {
    val map = new mutable.LinkedHashMap[String, java.util.List[String]]

    val fs = FileSystem.get(new Configuration())
    val path = new Path(filePath)
    if (fs.exists(path)) {
      val bufferedReader = new BufferedReader(new InputStreamReader(fs.open(path)))
      var line: String = null
      line = bufferedReader.readLine()
      while (line != null) {

        val fields: Array[String] = line.split(",")
        val int_id: String = fields(0)
        val date = new java.util.Date(java.lang.Long.valueOf(fields(2)))
        val time = sdf.format(date)
        System.out.println(line + "(" + time + ")")

        if (!map.keySet.contains(int_id))
          map.put(int_id, new java.util.ArrayList[String])
        map.get(int_id).get.add(time)

        line = bufferedReader.readLine()
      }

      map.toMap
    } else {
      throw new RuntimeException("the file do not exists")
    }
  }
}

测试日志：

18/11/19 16:50:15 INFO scheduler.DAGScheduler: Job 2 finished: print at App.scala:59, took 0.080061 s
-------------------------------------------
Time: 1542617415000 ms
-------------------------------------------
CellJoinMro(2,333,2018-11-05 00:00:00)
。。。。
18/11/19 16:50:15 INFO storage.BlockManagerInfo: Removed input-0-1542617392400 on 10.60.0.11:1337 in memory (size: 12.0 B, free: 456.1 MB)
》》》》》》》》》》》》》》》》此时路径上传新资源文件》》》》》》》》》》》》》》》》》》》》》》
更新broadcast...2018-11-19 16:50:30
。。。
1,111,1541433600000(2018-11-06 00:00:00)
2,222,1541433600000(2018-11-06 00:00:00)
3,333,1541433600000(2018-11-06 00:00:00)
18/11/19 16:50:30 INFO memory.MemoryStore: Block broadcast_5 stored as values in memory (estimated size 688.0 B, free 456.1 MB)
。。
18/11/19 16:50:30 INFO scheduler.JobScheduler: Starting job streaming job 1542617430000 ms.0 from job set of time 1542617430000 ms
-------------------------------------------
Time: 1542617430000 ms
-------------------------------------------

18/11/19 16:50:30 INFO scheduler.JobScheduler: Finished job streaming job 1542617430000 ms.0 from job set of time 1542617430000 ms
。。。。
18/11/19 16:50:32 WARN storage.BlockManager: Block input-0-1542617432400 replicated to only 0 peer(s) instead of 1 peers
18/11/19 16:50:32 INFO receiver.BlockGenerator: Pushed block input-0-1542617432400
更新broadcast...2018-11-19 16:50:45
1,111,1541433600000(2018-11-06 00:00:00)
2,222,1541433600000(2018-11-06 00:00:00)
3,333,1541433600000(2018-11-06 00:00:00)
18/11/19 16:50:45 INFO memory.MemoryStore: Block broadcast_6 stored as values in memory (estimated size 688.0 B, free 456.1 MB)
。。。。
18/11/19 16:50:45 INFO scheduler.DAGScheduler: Job 3 finished: print at App.scala:59, took 0.066975 s
-------------------------------------------
Time: 1542617445000 ms
-------------------------------------------
CellJoinMro(3,4444,2018-11-06 00:00:00)

18/11/19 16:50:45 INFO scheduler.JobScheduler: Finished job streaming job 1542617445000 ms.0 from job set of time 1542617445000 ms
18/11/19 16:50:45 INFO scheduler.JobScheduler: Total delay: 0.367 s for time 1542617445000 ms (execution: 0.083 s)
18/11/19 16:50:45 INFO rdd.MapPartitionsRDD: Removing RDD 9 from persistence list

日志分析：

每个batch都执行transform中的更新broadcast代码，而且也执行了broadcast获取代码。因此，每次都可进行更新broadcast内容，并且获取到broadcast中的内容。

2）Spark Structured Streaming更新broadcast（不可行）

 def sparkStructuredStreaming(): Unit = {
    val spark = SparkSession.builder.appName("Test_Broadcast_ByScala_App").getOrCreate()
    spark.streams.addListener(new StreamingQueryListener {
      override def onQueryStarted(event: StreamingQueryListener.QueryStartedEvent): Unit = {
        println("*************** onQueryStarted  ***************")
      }

      override def onQueryProgress(event: StreamingQueryListener.QueryProgressEvent): Unit = {
        println("*************** onQueryProgress  ***************")
        //        这段代码可以把broadcast对象更新成功，但是spark structured streaming内部读取到的broadcast对象数据依然是老数据。
        //        BroadcastWrapper.update(spark.sparkContext, true)
        println("*************** onQueryProgress update broadcast " + new SimpleDateFormat("yyyy-MM-dd HH:mm:ss").format(new java.util.Date()))

      }

      override def onQueryTerminated(event: StreamingQueryListener.QueryTerminatedEvent): Unit = {
        println("*************** onQueryTerminated  ***************")
      }
    })
    // Create DataFrame representing the stream of input lines from connection to localhost:9999
    val lines = spark.readStream.format("socket").option("host", ipAddr).option("port", 19999).load()

    import spark.implicits._
    val mro = lines.as(Encoders.STRING).map(row => {
      val fields = row.split(",")
      Mro(fields(0), fields(1))
    })

    val cellJoinMro = mro.transform(row => {
      //      这段代码在第一次触发时执行，之后触发就不再执行。
      println("更新broadcast..." + new SimpleDateFormat("yyyy-MM-dd HH:mm:ss").format(new java.util.Date()))
      if (1 < 3) {
        println("------------------------1111-----------------------------")
        BroadcastWrapper.update(spark.sparkContext)
      }
      var broadcastCellRes = BroadcastWrapper.getInstance(spark.sparkContext)
      row.map(row => {
        val int_id: String = row.int_id
        val rsrp: String = row.rsrp
        val findResult: String = String.join(",", broadcastCellRes.value.get(int_id).get)
        val timeStamps: String = String.join(",", findResult)

        CellJoinMro(int_id, rsrp, timeStamps)
      })
    })

    val query = cellJoinMro.writeStream.format("console")
      .outputMode("update")
      .trigger(Trigger.ProcessingTime(15, TimeUnit.SECONDS))
      .start()

    query.awaitTermination()
  }

执行日志：

18/11/19 17:12:49 INFO state.StateStoreCoordinatorRef: Registered StateStoreCoordinator endpoint
18/11/19 17:12:50 WARN streaming.TextSocketSourceProvider: The socket source should not be used for production applications! It does not support recovery.
更新broadcast...2018-11-19 17:12:51
------------------------1111-----------------------------
1,111,1541347200000(2018-11-05 00:00:00)
2,222,1541347200000(2018-11-05 00:00:00)
3,333,1541347200000(2018-11-05 00:00:00)
.....
-------------------------------------------
Batch: 0
-------------------------------------------
18/11/19 17:13:03 INFO codegen.CodeGenerator: Code generated in 82.760622 ms
。。。。
18/11/19 17:13:19 INFO scheduler.DAGScheduler: Job 4 finished: start at App.scala:109, took 4.215709 s
+------+----+-------------------+
|int_id|rsrp|          timestamp|
+------+----+-------------------+
|     1|  22|2018-11-05 00:00:00|
+------+----+-------------------+

18/11/19 17:14:00 INFO streaming.StreamExecution: Committed offsets for batch 1. Metadata OffsetSeqMetadata(0,1542618840003,Map(spark.sql.shuffle.partitions -> 600))

此时更新资源文件，附加2018-11-06的资源文件。
-------------------------------------------
Batch: 1
-------------------------------------------
18/11/19 17:14:00 INFO spark.SparkContext: Starting job: start at App.scala:109
。。。
18/11/19 17:14:05 INFO scheduler.DAGScheduler: Job 9 finished: start at App.scala:109, took 3.068106 s
+------+----+-------------------+
|int_id|rsrp|          timestamp|
+------+----+-------------------+
|     2| 333|2018-11-05 00:00:00|
+------+----+-------------------+

日志分析：

测试结论：

Spark Streaming更新broadcast（可行）、Spark Structured Streaming更新broadcast（不可行），原因Spark Streaming的执行引擎是Spark Engine，是代码执行，在算子的构造函数中可以访问SparkContext,SparkSession，而且这些类构造函数是可以每次都执行的。

而Spark Structured Streaming的执行引擎是Spark Sql Engine，是把代码优化为Spark Sql Engine希望的格式去执行，不可以在每次trigger事件触发都执行执行块以外的代码，因此这些类构造函数块代码只能执行一次，执行块类似MapFunction的call()函数内，不允许访问SparkContext,SparkSession对象，因此无处进行每次trigger都进行broadcast更新。

那么，如何在Spark Struectured Streaming中实现更新broadcast的方案，升级spark版本，从spark2.3.0开始，spark structured streaming支持了stream join stream（请参考《Spark2.3(三十七)：Stream join Stream（res文件每天更新一份）》）。