Spark Streaming 结合FlumeNG使用实例

SparkStreaming是一个对实时数据流进行高通量、容错处理的流式处理系统，可以对多种数据源（如Kdfka、Flume、Twitter、Zero和TCP 套接字）进行类似map、reduce、join、window等复杂操作，并将结果保存到外部文件系统、数据库或应用到实时仪表盘。

Spark Streaming流式处理系统特点有：

• 将流式计算分解成一系列短小的批处理作业
• 将失败或者执行较慢的任务在其它节点上并行执行
• 较强的容错能力(基于RDD继承关系Lineage)
• 使用和RDD一样的语义

本文将Spark Streaming结合FlumeNG，然后以源码中的JavaFlumeEventCount作参考，建立maven工程，打包在spark standalone集群运行。

一、步骤

1.建立maven工程，写好pom.xml

需要spark streaming的flume插件包，jar的maven地址如下，填入pom.xml中

 <dependency>
     <groupId>org.apache.spark</groupId>
     <artifactId>spark-streaming-flume_2.10</artifactId>
     <version>1.1.0</version>
 </dependency>

完整的pom.xml

<project xmlns="http://maven.apache.org/POM/4.0.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
    xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
    <modelVersion>4.0.0</modelVersion>
    <groupId>test</groupId>
    <artifactId>hq</artifactId>
    <version>0.0.1-SNAPSHOT</version>
    <build>
    <plugins>
            <plugin>
                <groupId>org.apache.maven.plugins</groupId>
                <artifactId>maven-compiler-plugin</artifactId>
                <version>2.3.2</version>
                <configuration>
                    <source>1.6</source>
                    <target>1.6</target>
                    <compilerVersion>1.6</compilerVersion>
                    <encoding>UTF-8</encoding>
                </configuration>
            </plugin>
            <plugin>
                <groupId>org.apache.maven.plugins</groupId>
                <artifactId>maven-jar-plugin</artifactId>
                <version>2.3.2</version>
                <configuration>
                    <archive>
                        <manifest>
                            <addClasspath>true</addClasspath>
                            <classpathPrefix>.</classpathPrefix>
                            <mainClass>JavaFlumeEventCount</mainClass>
                        </manifest>
                    </archive>
                </configuration>
            </plugin>
            <plugin>
                <groupId>org.apache.maven.plugins</groupId>
                <artifactId>maven-assembly-plugin</artifactId>
                <version>2.4</version>
                <configuration>
                  <descriptorRefs>
                    <descriptorRef>jar-with-dependencies</descriptorRef>
                  </descriptorRefs>
                </configuration>
            </plugin>
        </plugins>
    </build>
    <dependencies>
        <dependency>
            <groupId>org.apache.spark</groupId>
            <artifactId>spark-streaming-flume_2.10</artifactId>
            <version>1.1.0</version>
        </dependency>
    </dependencies>
</project>

2.编码并且打包

JavaCode：

import org.apache.spark.SparkConf;
import org.apache.spark.api.java.function.Function;
import org.apache.spark.streaming.*;
import org.apache.spark.streaming.api.java.*;
import org.apache.spark.streaming.flume.FlumeUtils;
import org.apache.spark.streaming.flume.SparkFlumeEvent;

public final class JavaFlumeEventCount {
    private JavaFlumeEventCount() {
    }

    public static void main(String[] args) {

        String host = args[0];
        int port = Integer.parseInt(args[1]);

        Duration batchInterval = new Duration(Integer.parseInt(args[2]));
        SparkConf sparkConf = new SparkConf().setAppName("JavaFlumeEventCount");
        JavaStreamingContext ssc = new JavaStreamingContext(sparkConf,
                batchInterval);
        JavaReceiverInputDStream<SparkFlumeEvent> flumeStream = FlumeUtils
                .createStream(ssc, host, port);

        flumeStream.count();

        flumeStream.count().map(new Function<Long, String>() {
            private static final long serialVersionUID = -572435064083746235L;

            public String call(Long in) {
                return "Received " + in + " flume events.";
            }
        }).print();

        ssc.start();
        ssc.awaitTermination();
    }
}

maven 命令：eclipse中run as -> Maven Assembly:assembly

得到工程的target目录下得到jar包：hq-0.0.1-SNAPSHOT.jar

3.将3个jar包上传到服务器，准备运行

除了自身打的jar包外，运行还需要：spark-streaming-flume_2.10-1.1.0.jar,flume-ng-sdk-1.4.0.jar 这2个jar包（我使用的flume-ng版本是1.4.0）

将3个jar包上传到服务器~/spark/test/目录下。

4.命令行提交任务，运行

[ebupt@eb174 test]$ spark-submit --master spark://eb174:7077 --name FlumeStreaming --class JavaFlumeEventCount --executor-memory 1G --total-executor-cores 2 --jars spark-streaming-flume_2.10-1.1.0.jar,flume-ng-sdk-1.4.0.jar hq.jar eb174 11000 5000

注意：参数解释：spark-submit --help。自己可以根据需要修改内存，防止OOM。另外jars可以同时加载多个jar包，逗号分隔。指定的运行类后需要指定3个参数。

5.开启flume-ng，启动数据源

书写好flume的agent配置文件spark-flumeng.conf，内容如下：

 #Agent5
 #List the sources, sinks and channels for the agent
 agent5.sources =  source1
 agent5.sinks =  hdfs01
 agent5.channels = channel1

 #set channel for sources and sinks
 agent5.sources.source1.channels = channel1
 agent5.sinks.hdfs01.channel = channel1

 #properties of someone source
 agent5.sources.source1.type = spooldir
 agent5.sources.source1.spoolDir = /home/hadoop/huangq/spark-flumeng-data/
 agent5.sources.source1.ignorePattern = .*(\\.index|\\.tmp|\\.xml)$
 agent5.sources.source1.fileSuffix = .1
 agent5.sources.source1.fileHeader = true
 agent5.sources.source1.fileHeaderKey = filename

 # set interceptors
 agent5.sources.source1.interceptors = i1 i2
 agent5.sources.source1.interceptors.i1.type = org.apache.flume.interceptor.HostInterceptor$Builder
 agent5.sources.source1.interceptors.i1.preserveExisting = false
 agent5.sources.source1.interceptors.i1.hostHeader = hostname
 agent5.sources.source1.interceptors.i1.useIP=false
 agent5.sources.source1.interceptors.i2.type = org.apache.flume.interceptor.TimestampInterceptor$Builder

 #properties of mem-channel-1
 agent5.channels.channel1.type = memory
 agent5.channels.channel1.capacity = 100000
 agent5.channels.channel1.transactionCapacity = 100000
 agent5.channels.channel1.keep-alive = 30

 #properties of sink
 agent5.sinks.hdfs01.type = avro
 agent5.sinks.hdfs01.hostname = eb174
 agent5.sinks.hdfs01.port = 11000

启动flume-ng: [hadoop@eb170 flume]$ bin/flume-ng agent -n agent5 -c conf  -f conf/spark-flumeng.conf

注意：

①flume的sink要用avro，指定要发送到的spark集群中的一个节点，我们这里是eb174:11000。

②如果没有指定Flume的sdk包，会出现错误：　java.lang.NoClassDefFoundError: Lorg/apache/flume/source/avro/AvroFlumeEvent;没有找到类。这个类在flume的sdk包内，在jars参数中指定jar包位置就可以。

③将自己定义的运行jar包单独列出，不要放在jars参数指定，否则也会有错误抛出。

6.运行结果

在提交spark任务的客户端可以看到，看到大量的输出信息，然后可以看到有数据的RDD会统计出这个RDD有多少行，统计结果如下：

 Spark assembly has been built with Hive, including Datanucleus jars on classpath
 Using Spark's default log4j profile: org/apache/spark/log4j-defaults.properties
 14/10/13 19:00:44 INFO SecurityManager: Changing view acls to: ebupt,
 14/10/13 19:00:44 INFO SecurityManager: Changing modify acls to: ebupt,
 14/10/13 19:00:44 INFO SecurityManager: SecurityManager: authentication disabled; ui acls disabled; users with view permissions: Set(ebupt, ); users with modify permissions: Set(ebupt, )
 14/10/13 19:00:45 INFO Slf4jLogger: Slf4jLogger started
 14/10/13 19:00:45 INFO Remoting: Starting remoting
 14/10/13 19:00:45 INFO Remoting: Remoting started; listening on addresses :[akka.tcp://sparkDriver@eb174:51147]
 14/10/13 19:00:45 INFO Remoting: Remoting now listens on addresses: [akka.tcp://sparkDriver@eb174:51147]
 14/10/13 19:00:45 INFO Utils: Successfully started service 'sparkDriver' on port 51147.
 14/10/13 19:00:45 INFO SparkEnv: Registering MapOutputTracker
 14/10/13 19:00:45 INFO SparkEnv: Registering BlockManagerMaster
 ....
 .....
 14/10/13 19:09:21 INFO DAGScheduler: Missing parents: List()
 14/10/13 19:09:21 INFO DAGScheduler: Submitting Stage 145 (MappedRDD[291] at map at MappedDStream.scala:35), which has no missing parents
 14/10/13 19:09:21 INFO MemoryStore: ensureFreeSpace(3400) called with curMem=13047, maxMem=278302556
 14/10/13 19:09:21 INFO MemoryStore: Block broadcast_110 stored as values in memory (estimated size 3.3 KB, free 265.4 MB)
 14/10/13 19:09:21 INFO MemoryStore: ensureFreeSpace(2020) called with curMem=16447, maxMem=278302556
 14/10/13 19:09:21 INFO MemoryStore: Block broadcast_110_piece0 stored as bytes in memory (estimated size 2020.0 B, free 265.4 MB)
 14/10/13 19:09:21 INFO BlockManagerInfo: Added broadcast_110_piece0 in memory on eb174:41187 (size: 2020.0 B, free: 265.4 MB)
 14/10/13 19:09:21 INFO BlockManagerMaster: Updated info of block broadcast_110_piece0
 14/10/13 19:09:21 INFO DAGScheduler: Submitting 1 missing tasks from Stage 145 (MappedRDD[291] at map at MappedDStream.scala:35)
 14/10/13 19:09:21 INFO TaskSchedulerImpl: Adding task set 145.0 with 1 tasks
 14/10/13 19:09:21 INFO TaskSetManager: Starting task 0.0 in stage 145.0 (TID 190, eb175, PROCESS_LOCAL, 1132 bytes)
 14/10/13 19:09:21 INFO BlockManagerInfo: Added broadcast_110_piece0 in memory on eb175:57696 (size: 2020.0 B, free: 519.6 MB)
 14/10/13 19:09:21 INFO TaskSetManager: Finished task 0.0 in stage 145.0 (TID 190) in 25 ms on eb175 (1/1)
 14/10/13 19:09:21 INFO DAGScheduler: Stage 145 (take at DStream.scala:608) finished in 0.026 s
 14/10/13 19:09:21 INFO TaskSchedulerImpl: Removed TaskSet 145.0, whose tasks have all completed, from pool
 14/10/13 19:09:21 INFO SparkContext: Job finished: take at DStream.scala:608, took 0.036589357 s
 -------------------------------------------
 Time: 1413198560000 ms
 -------------------------------------------
 Received 35300 flume events.

 14/10/13 19:09:55 INFO JobScheduler: Finished job streaming job 1413198595000 ms.0 from job set of time 1413198595000 ms
 14/10/13 19:09:55 INFO JobScheduler: Total delay: 0.126 s for time 1413198595000 ms (execution: 0.112 s)
 14/10/13 19:09:55 INFO MappedRDD: Removing RDD 339 from persistence list
 14/10/13 19:09:55 INFO BlockManager: Removing RDD 339
 14/10/13 19:09:55 INFO MappedRDD: Removing RDD 338 from persistence list
 14/10/13 19:09:55 INFO BlockManager: Removing RDD 338
 14/10/13 19:09:55 INFO MappedRDD: Removing RDD 337 from persistence list
 14/10/13 19:09:55 INFO BlockManager: Removing RDD 337
 14/10/13 19:09:55 INFO ShuffledRDD: Removing RDD 336 from persistence list
 14/10/13 19:09:55 INFO BlockManager: Removing RDD 336
 14/10/13 19:09:55 INFO UnionRDD: Removing RDD 335 from persistence list
 14/10/13 19:09:55 INFO BlockManager: Removing RDD 335
 14/10/13 19:09:55 INFO MappedRDD: Removing RDD 333 from persistence list
 14/10/13 19:09:55 INFO BlockManager: Removing RDD 333
 14/10/13 19:09:55 INFO BlockRDD: Removing RDD 332 from persistence list
 14/10/13 19:09:55 INFO BlockManager: Removing RDD 332
 ...
 ...
 14/10/13 19:10:00 INFO TaskSchedulerImpl: Adding task set 177.0 with 1 tasks
 14/10/13 19:10:00 INFO TaskSetManager: Starting task 0.0 in stage 177.0 (TID 215, eb175, PROCESS_LOCAL, 1132 bytes)
 14/10/13 19:10:00 INFO BlockManagerInfo: Added broadcast_134_piece0 in memory on eb175:57696 (size: 2021.0 B, free: 530.2 MB)
 14/10/13 19:10:00 INFO TaskSetManager: Finished task 0.0 in stage 177.0 (TID 215) in 24 ms on eb175 (1/1)
 14/10/13 19:10:00 INFO DAGScheduler: Stage 177 (take at DStream.scala:608) finished in 0.024 s
 14/10/13 19:10:00 INFO TaskSchedulerImpl: Removed TaskSet 177.0, whose tasks have all completed, from pool
 14/10/13 19:10:00 INFO SparkContext: Job finished: take at DStream.scala:608, took 0.033844743 s
 -------------------------------------------
 Time: 1413198600000 ms
 -------------------------------------------
 Received 0 flume events.

二、结论

• flume-ng与spark的结合成功，可根据需要灵活编写相关的类来实现实时处理FlumeNG传输的数据。
• spark streaming和多种数据源结合，达到实时计算处理的能力。

三、参考资料

1. Spark Streaming和Flume-NG对接实验
2. Spark和Flume-ng整合
3. Flume sink 配置手册