前言
最近在利用Spark streaming和Kafka构建一个实时的数据分析系统，对图书阅读数据进行分析，做实时推荐。Spark Streaming 模块是对于 Spark Core 的一个扩展，目的是为了以高吞吐量，并且容错的方式处理持续性的数据流。目前 Spark Streaming 支持的外部数据源有 Flume、 Kafka、Twitter、ZeroMQ、TCP Socket 等。Apache Kafka是一个分布式的消息发布-订阅系统，Kafka可以作为流计算系统的数据源，本例中Spark Streaming将从Kafka中消费数据。
系统环境
软件版本
1
2
3 |
Spark: 1.4.1
Kafka: 0.8.1.1
zookeeper: 3.4.6 |
集群节点
一共有四台主机，主机名分别为nn0001, dn0001, dn0002, dn0003。
1
2
3
4 |
192.168.186.12 nn0001
192.168.186.13 dn0001
192.168.186.14 dn0002
192.168.186.15 dn0003 |
zookeeper安装
kafka使用zookeeper来管理，存储一些meta信息，并使用了zookeeper watch机制来发现meta信息的变更并作出相应的动作(比如consumer失效,触发负载均衡等)。
Zookeeper的配置在机器1上完成后分发到其他三台机器即可。
1
2
3
4
5
6 |
[bigdata@nn0001 ~]$ wget http:
[bigdata@nn0001 ~]$ tar -zxvf zookeeper-3.4.6.tar.gz
[bigdata@nn0001 ~]$cd zookeeper-3.4.6/conf
[bigdata@nn0001 conf]$ pwd
/home/bigdata/bigprosoft/zookeeper-3.4.6/conf
[bigdata@nn0001 conf]$ cp zoo_sample.cfg zoo.cfg |
修改配置文件
1
2
3
4
5
6
7
8
9
10 |
[bigdata@nn0001 conf]$ vi zoo.cfg
tickTime=2000
dataDir=/home/bigdata/bigprosoft/zookeeper/data
clientPort=2181
initLimit=10
syncLimit=5
server.1=nn0001:2888:3888
server.2=dn0001:2888:3888
server.3=dn0002:2888:3888
server.4=dn0003:2888:3888 |
在dataDir目录下创建myid文件，nn0001机器的内容为1，dn0001机器的内容为2，更多依此类推。
1
2
3 |
[bigdata@nn0001 data]$ echo 1 > myid
[bigdata@nn0001 data]$ cat myid
1 |
启动测试
1
2
3
4
5
6
7
8
9
10 |
[bigdata@nn0001 bin]$ ./zkServer.sh start
[bigdata@nn0001 bin]$ jps
10805 QuorumPeerMain #已经启动成功了
15494 Master
11816 NameNode
20958 Jps
17539 Worker
12084 ResourceManager
12945 RunJar
12944 RunJar |
停止
1 |
[bigdata@nn0001 bin]$ ./zkServer.sh stop |
其它机器相同操作,scp过去即可。
kafka安装
Kafka的broker、producer、consumer、topic等概念以及原理可以查阅官方文档
本次实验采用的多节点多broker集群模式，为每一台机器分配一个broker id。
1
2
3
4
5
6
7
8
9
10 |
[bigdata@nn0001 ~]$ wget http:
[bigdata@nn0001 ~]$ tar zxf kafka_2.10-0.8.1.1.tgz
[bigdata@nn0001 ~]$ cd kafka_2.10-0.8.1.1
[bigdata@nn0001 kafka_2.10-0.8.1.1]$ cd conf
[bigdata@nn0001 conf]$ vi server.properties
broker.id=1 #其它机器的id依次递增即可
port=9092
host.name=192.168.186.12
advertised.host.name=192.168.186.12
zookeeper.connect=192.168.186.12:2181,192.168.186.13:2181,192.168.186.14:2181,192.168.186.15:2181 |
修改完成后分发到另外三台机器上。
启动测试
1
2
3
4
5
6
7
8
9
10
11 |
[bigdata@nn0001 bin]$ nohup ./kafka-server-start.sh ../config/server.properties &
[bigdata@nn0001 conf]$ jps
10805 QuorumPeerMain
21282 Jps
15494 Master
21209 Kafka
11816 NameNode
17539 Worker
12084 ResourceManager
12945 RunJar
12944 RunJar |
依次启动机器
kafka使用测试
创建topic
1 |
[ ] .-. -- -- --- -- -- |
查看topic
1
2
3
4
5
6 |
[bigdata@nn0001 bin]$ ./kafka-topics.sh --describe --zookeeper nn0001:2181
Topic:mytest PartitionCount:2 ReplicationFactor:2 Configs:
Topic: mytest Partition: 0 Leader: 2 Replicas: 3,2 Isr: 2
Topic: mytest Partition: 1 Leader: -1 Replicas: 4,3 Isr:
Topic:test PartitionCount:1 ReplicationFactor:3 Configs:
Topic: test Partition: 0 Leader: 2 Replicas: 2,3,4 Isr: 2 |
producer测试
1
2
3 |
[bigdata@nn0001 bin]$ ./kafka-console-producer.sh --broker-list 192.168.186.12:9092 --topic test
gsdggfgfgfd
gdfgdfgdf |
conumer测试
1
2
3
4
5
6
7
8
9
10 |
[bigdata@nn0001 bin]$ ./kafka-console-consumer.sh --zookeeper 192.168.186.12:2181 --from-beginning --topic test
abfsfsdfsdfs
ffsdfs
gsdggfgfgfd
gdfgdfgdf
^C[2015-08-28 17:48:40,991] WARN Reconnect due to socket error: null (kafka.consumer.SimpleConsumer)
Consumed 7 messages
` |
测试高可用
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22 |
[bigdata@nn0001 bin]$ ./kafka-topics.sh --describe --zookeeper 192.168.186.12:2181,192.168.186.13:2181,192.168.186.14:2181,192.168.186.15:2181 --from-beginning --topic test
Topic:test PartitionCount:1 ReplicationFactor:3 Configs:
Topic: test Partition: 0 Leader: 2 Replicas: 2,3,4 Isr: 2,4
#可以看到leader是2，是dn0001机器，把此机器上的kafka进程杀掉，再查看topic的leader
[bigdata@dn0002 bin]$ ./kafka-topics.sh --describe --zookeeper 192.168.186.12:2181,192.168.186.13:2181,192.168.186.14:2181,192.168.186.15:2181 --topic test
Topic:test PartitionCount:1 ReplicationFactor:3 Configs:
Topic: test Partition: 0 Leader: 4 Replicas: 2,3,4 Isr: 4
#此时leader变成了4，对应的机器是dn0003.
[bigdata@nn0001 bin]$ ./kafka-console-consumer.sh --zookeeper 192.168.186.12:2181,192.168.186.13:2181,192.168.186.14:2181,192.168.186.15:2181 --from-beginning --topic test
abfsfsdfsdfs
ffsdfs
gsdggfgfgfd
gdfgdfgdf
q
^C[2015-08-31 10:14:50,964] WARN Reconnect due to socket error: null (kafka.consumer.SimpleConsumer)
Consumed 7 messages
#消费者消费信息测试 |
ok，搭建过程就完成，下面用python/java/scala进行开发实例即可。
排错
问题1描述
1
2
3 |
SLF4J: Failed to load class "org.slf4j.impl.StaticLoggerBinder".
SLF4J: Defaulting to no-operation (NOP) logger implementation
SLF4J: See http://www.slf4j.org/codes.html#StaticLoggerBinder for further details. |
解决方法
1
2
3 |
[bigdata@nn0001 ~]$ wget http:
[bigdata@nn0001 ~]$ cd slf4j-1.7.12
[bigdata@nn0001 ~]$ cp slf4j-nop-1.7.12.jar ~/bigprosoft/kafka/libs/ |
问题2描述
1
2
3
4
5
6
7
8
9
10
11
12
13
14 |
[bigdata@nn0001 bin]$ ./kafka-console-producer.sh --broker-list nn0001:9092 --topic test
fsfsdfsdf
……
[2015-08-28 17:24:18,417] ERROR Failed to send requests for topics test with correlation ids in [0,8] (kafka.producer.async.DefaultEventHandler)
[2015-08-28 17:24:18,419] ERROR Error in handling batch of 1 events (kafka.producer.async.ProducerSendThread)
kafka.common.FailedToSendMessageException: Failed to send messages after 3 tries.
at kafka.producer.async.DefaultEventHandler.handle(DefaultEventHandler.scala:90)
at kafka.producer.async.ProducerSendThread.tryToHandle(ProducerSendThread.scala:104)
at kafka.producer.async.ProducerSendThread$$anonfun$processEvents$3.apply(ProducerSendThread.scala:87)
at kafka.producer.async.ProducerSendThread$$anonfun$processEvents$3.apply(ProducerSendThread.scala:67)
at scala.collection.immutable.Stream.foreach(Stream.scala:547)
at kafka.producer.async.ProducerSendThread.processEvents(ProducerSendThread.scala:66)
at kafka.producer.async.ProducerSendThread.run(ProducerSendThread.scala:44)
…… |
解决方法，把server.properties中主机名改为IP地址即可。
1
2
3 |
host.name=10.171.59.221
advertised.host.name=10.171.59.221
zookeeper.connect=192.168.186.12:2181,192.168.186.13:2181,192.168.186.14:2181,192.168.186.15:2181 |
Kafka 的详细介绍：请点这里
Kafka 的下载地址：请点这里