Spark Streaming揭秘 Day28

在集成开发环境中详解Spark Streaming的运行日志内幕

今天会逐行解析一下SparkStreaming运行的日志，运行的是WordCountOnline这个Demo。

启动过程

SparkStreaming启动是从如下日志开始：

16/06/16 21:26:44 INFO ReceiverTracker: Starting 1 receivers
16/06/16 21:26:44 INFO ReceiverTracker: ReceiverTracker started
16/06/16 21:26:44 INFO ReceiverTracker: Receiver 0 started

ReceiverTracker启动了一个receiver，完成启动

16/06/16 21:26:44 INFO ForEachDStream: metadataCleanupDelay = -1
16/06/16 21:26:44 INFO ShuffledDStream: metadataCleanupDelay = -1
16/06/16 21:26:44 INFO MappedDStream: metadataCleanupDelay = -1
16/06/16 21:26:44 INFO FlatMappedDStream: metadataCleanupDelay = -1
16/06/16 21:26:44 INFO SocketInputDStream: metadataCleanupDelay = -1

从日志的顺序，可以看出从Driver的角度讲，DStreamGraph是从后往前构造的，一直到最开始的地方SocketInputDStream

16/06/16 21:26:44 INFO SocketInputDStream: Slide time = 15000 ms
16/06/16 21:26:44 INFO SocketInputDStream: Storage level = StorageLevel(false, false, false, false, 1)
16/06/16 21:26:44 INFO SocketInputDStream: Checkpoint interval = null
16/06/16 21:26:44 INFO SocketInputDStream: Remember duration = 15000 ms
16/06/16 21:26:44 INFO SocketInputDStream: Initialized and validated org.apache.spark.streaming.dstream.SocketInputDStream@2d14d3a1
....
16/06/16 21:26:44 INFO ForEachDStream: Slide time = 15000 ms
16/06/16 21:26:44 INFO ForEachDStream: Storage level = StorageLevel(false, false, false, false, 1)
16/06/16 21:26:44 INFO ForEachDStream: Checkpoint interval = null
16/06/16 21:26:44 INFO ForEachDStream: Remember duration = 15000 ms
16/06/16 21:26:44 INFO ForEachDStream: Initialized and validated org.apache.spark.streaming.dstream.ForEachDStream@57481f9

这部分日志是在SparkStreaming运行前的实例化过程，构建起DStreamGraph中的对象。
Slide time和Remember duration分别表示Batch Duration和Remember Duration。
StorageLevel部分的日志有点问题，从源码看应该是StorageLevel(true, true, false, false, 2)。

16/06/16 21:26:44 INFO RecurringTimer: Started timer for JobGenerator at time 1466083605000
16/06/16 21:26:44 INFO JobGenerator: Started JobGenerator at 1466083605000 ms
16/06/16 21:26:44 INFO JobScheduler: Started JobScheduler
16/06/16 21:26:44 INFO StreamingContext: StreamingContext started

这里启动了定时器支持JobGenerator工作，JobGenerator与定期器结合能不断的产生Job，
定时器产生Job之后，放在线程中，通过JobScheduler提交给SparkCore来进行运行。

简单总结下：

在启动时会在Executor上启动receivers。
在这个基础上，因为要构造出DStreamGraph，会从后往前回溯链条，并从前往后构造对象。
RecurringTimer、JobGenerator、JobScheduler是调度的核心。

至此，Driver和Receiver都启动完成。

接收数据

接下来是接收数据，定时器会根据我们设定的时间开始工作，这里是关键。

16/06/16 21:26:44 INFO RecurringTimer: Started timer for BlockGenerator at time 1466083605000
16/06/16 21:26:44 INFO BlockGenerator: Started BlockGenerator
16/06/16 21:26:44 INFO BlockGenerator: Started block pushing thread
启动线程存数据
16/06/16 21:26:44 INFO ReceiverTracker: Registered receiver for stream 0 from 192.168.1.120:51324
对Receiver进行注册

下面开始在Executor上执行，启动Receiver
16/06/16 21:26:44 INFO ReceiverSupervisorImpl: Starting receiver
16/06/16 21:26:44 INFO ReceiverSupervisorImpl: Called receiver onStart
16/06/16 21:26:44 INFO ReceiverSupervisorImpl: Waiting for receiver to be stopped
16/06/16 21:26:44 INFO SocketReceiver: Connecting to localhost:9999
16/06/16 21:26:44 INFO SocketReceiver: Connected to localhost:9999

下面是具体存储数据
16/06/16 21:26:45 INFO MemoryStore: Block input-0-1466083604800 stored as bytes in memory (estimated size 10.0 B, free 45.1 KB)
16/06/16 21:26:45 INFO BlockManagerInfo: Added input-0-1466083604800 in memory on localhost:51326 (size: 10.0 B, free: 1140.4 MB)
16/06/16 21:26:45 WARN BlockManager: Block input-0-1466083604800 replicated to only 0 peer(s) instead of 1 peers
16/06/16 21:26:45 INFO BlockGenerator: Pushed block input-0-1466083604800

下面回到Driver，产生了Job
16/06/16 21:26:45 INFO JobScheduler: Added jobs for time 1466083605000 ms
16/06/16 21:26:45 INFO JobScheduler: Starting job streaming job 1466083605000 ms.0 from job set of time 1466083605000 ms

再看下执行完成后日志：
16/06/16 21:26:45 INFO JobScheduler: Finished job streaming job 1466083605000 ms.0 from job set of time 1466083605000 ms
16/06/16 21:26:45 INFO JobScheduler: Total delay: 0.188 s for time 1466083605000 ms (execution: 0.127 s)
这里说明了延时时间。

清理数据

16/06/16 21:26:45 INFO ReceivedBlockTracker: Deleting batches ArrayBuffer()
16/06/16 21:26:45 INFO InputInfoTracker: remove old batch metadata:

之后是删除数据，我们发现，执行清理RDD的动作是有点滞后的，差了一个Batch Duration，也就是说在下一个Batch Duration的时候，清理前一次的。
16/06/16 21:27:00 INFO ShuffledRDD: Removing RDD 4 from persistence list
16/06/16 21:27:00 INFO MapPartitionsRDD: Removing RDD 3 from persistence list
16/06/16 21:27:00 INFO MapPartitionsRDD: Removing RDD 2 from persistence list
16/06/16 21:27:00 INFO BlockRDD: Removing RDD 1 from persistence list
16/06/16 21:27:00 INFO BlockManager: Removing RDD 4
16/06/16 21:27:00 INFO BlockManager: Removing RDD 3
16/06/16 21:27:00 INFO BlockManager: Removing RDD 2
16/06/16 21:27:00 INFO BlockManager: Removing RDD 1
16/06/16 21:27:00 INFO SocketInputDStream: Removing blocks of RDD BlockRDD[1] at socketTextStream at WordCountOnline.java:58 of time 1466083620000 ms
完成数据删除后，是删除元数据。
16/06/16 21:27:00 INFO ReceivedBlockTracker: Deleting batches ArrayBuffer()
16/06/16 21:27:00 INFO InputInfoTracker: remove old batch metadata:

想补充一点的是，这个清理策略会和rememberDuration(window操作时使用)的设置有关，而rememberDuration默认和BatchDuration是一致的，代码也可以印证这点。

欲知后事如何，且听下回分解!

DT大数据每天晚上20：00YY频道现场授课频道68917580