Kafka实战：从RDBMS到Hadoop，七步实现实时传输

原文链接：Kafka in Action: 7 Steps to Real-Time Streaming From RDBMS to Hadoop
作者：Rajesh Nadipalli
译者：郭芮，关注大数据领域，寻求报道或投稿请联系guorui@csdn.net。

本文是关于Flume成功应用Kafka的研究案例，深入剖析它是如何将RDBMS实时数据流导入到HDFS的Hive表中。

对于那些想要把数据快速摄取到Hadoop中的企业来讲，Kafka是一个很好的选择。Kafka是什么？Kafka是一个分布式、可伸缩、可信赖的消息传递系统，利用发布-订阅模型来集成应用程序/数据流。同时，Kafka还是Hadoop技术堆栈中的关键组件，能够很好地支持实时数据分析或者货币化的物联网数据。

本文服务于技术人群。下面就图解Kafka是如何把数据流从RDBMS（关系数据库管理系统）导入Hive，并且成功支持了一个实时分析案例。作为参考，本文中使用的组件版本分别为Hive 1.2.1，Flume 1.6 以及 Kafka 0.9。

Kafka所在位置：解决方案的整体结构

下图显示了解决方案的整体结构：Kafka和Flume的结合，再加上Hive的事务特性，RDBMS的事务数据被成功传递到目标Hive表中。

七步实现Hadoop实时数据导入

现在让我们深入方案细节，并展示如何在几个步骤内将数据流导入Hadoop。

1.从RDBMS中提取数据

所有关系型数据库都有一个日志文件，用来记录最新的事务。解决方案的第一步就是获取这些事务数据，同时要确保这些数据格式是可以被Hadoop所接受的。

2.设置Kafka消息发送端

发布Kafka话题消息的进程称为“消息发送端”。“话题”里有各种Kafka所需要维护的信息类别，RDBMS数据也会被转换成Kafka话题。对于这个示例，要求设置一个服务于整个销售团队的数据库，且该数据库中的事务数据均以Kafka话题形式发布。以下步骤都需要设置Kafka消息发送端：

$ cd /usr/hdp/2.4.0.0-169/kafka
$ bin/kafka-topics.sh --create --zookeeper sandbox.hortonworks.com:2181 --replication-factor 1 --partitions 1 --topic SalesDBTransactions
Created topic "SalesDBTransactions".
$ bin/kafka-topics.sh --list --zookeeper sandbox.hortonworks.com:2181
SalesDBTransactions

3.设置Hive

接下来将创建一个Hive表，准备接收销售团队的数据库事务。这个例子中，我们将创建一个用户数据表：

[bedrock@sandbox ~]$ beeline -u jdbc:hive2:// -n hive -p hive
0: jdbc:hive2://> use raj;
create table customers (id string, name string, email string, street_address string, company string)
partitioned by (time string)
clustered by (id) into 5 buckets stored as orc
location '/user/bedrock/salescust'
TBLPROPERTIES ('transactional'='true');

为了确保Hive能够有效处理事务数据，以下设置要求在Hive配置中进行：

hive.txn.manager = org.apache.hadoop.hive.ql.lockmgr.DbTxnManager 

4.为Kafka到Hive的数据流设置Flume代理

现在来看下如何创建一个Flume代理，用于收集Kafka话题资料并向Hive表发送数据。

在启用Flume代理前，要通过这几个步骤设置运行环境：

$ pwd
/home/bedrock/streamingdemo
$ mkdir flume/checkpoint
$ mkdir flume/data
$ chmod 777 -R flume
$ export HIVE_HOME=/usr/hdp/current/hive-server2
$ export HCAT_HOME=/usr/hdp/current/hive-webhcat

$ pwd
/home/bedrock/streamingdemo/flume
$ mkdir logs

再如下所示创建一个log4j属性文件：

[bedrock@sandbox conf]$ vi log4j.properties
flume.root.logger=INFO,LOGFILE
flume.log.dir=/home/bedrock/streamingdemo/flume/logs
flume.log.file=flume.log

然后为Flume代理配置以下文件：

$ vi flumetohive.conf
flumeagent1.sources = source_from_kafka
flumeagent1.channels = mem_channel
flumeagent1.sinks = hive_sink
# Define / Configure source
flumeagent1.sources.source_from_kafka.type = org.apache.flume.source.kafka.KafkaSource
flumeagent1.sources.source_from_kafka.zookeeperConnect = sandbox.hortonworks.com:2181
flumeagent1.sources.source_from_kafka.topic = SalesDBTransactions
flumeagent1.sources.source_from_kafka.groupID = flume
flumeagent1.sources.source_from_kafka.channels = mem_channel
flumeagent1.sources.source_from_kafka.interceptors = i1
flumeagent1.sources.source_from_kafka.interceptors.i1.type = timestamp
flumeagent1.sources.source_from_kafka.consumer.timeout.ms = 1000

# Hive Sink
flumeagent1.sinks.hive_sink.type = hive
flumeagent1.sinks.hive_sink.hive.metastore = thrift://sandbox.hortonworks.com:9083
flumeagent1.sinks.hive_sink.hive.database = raj
flumeagent1.sinks.hive_sink.hive.table = customers
flumeagent1.sinks.hive_sink.hive.txnsPerBatchAsk = 2
flumeagent1.sinks.hive_sink.hive.partition = %y-%m-%d-%H-%M
flumeagent1.sinks.hive_sink.batchSize = 10
flumeagent1.sinks.hive_sink.serializer = DELIMITED
flumeagent1.sinks.hive_sink.serializer.delimiter = ,
flumeagent1.sinks.hive_sink.serializer.fieldnames = id,name,email,street_address,company
# Use a channel which buffers events in memory
flumeagent1.channels.mem_channel.type = memory
flumeagent1.channels.mem_channel.capacity = 10000
flumeagent1.channels.mem_channel.transactionCapacity = 100
# Bind the source and sink to the channel
flumeagent1.sources.source_from_kafka.channels = mem_channel
flumeagent1.sinks.hive_sink.channel = mem_channel

5.启用Flume代理

通过以下指令启用Flume代理：

$ /usr/hdp/apache-flume-1.6.0/bin/flume-ng agent -n flumeagent1 -f ~/streamingdemo/flume/conf/flumetohive.conf 

6.启用Kafka流

作为示例下面是一个模拟事务的消息集，这在实际系统中需要通过源数据库才能生成。例如，以下可能来自Oracle流，在回放被提交到数据库的SQL事务，也可能来自GoldenGate。

$ cd /usr/hdp/2.4.0.0-169/kafka
$ bin/kafka-console-producer.sh --broker-list sandbox.hortonworks.com:6667 --topic SalesDBTransactions
1,"Nero Morris","porttitor.interdum@Sedcongue.edu","P.O. Box 871, 5313 Quis Ave","Sodales Company"
2,"Cody Bond","ante.lectus.convallis@antebibendumullamcorper.ca","232-513 Molestie Road","Aenean Eget Magna Incorporated"
3,"Holmes Cannon","a@metusAliquam.edu","P.O. Box 726, 7682 Bibendum Rd.","Velit Cras LLP"
4,"Alexander Lewis","risus@urna.edu","Ap #375-9675 Lacus Av.","Ut Aliquam Iaculis Inc."
5,"Gavin Ortiz","sit.amet@aliquameu.net","Ap #453-1440 Urna. St.","Libero Nec Ltd"
6,"Ralph Fleming","sociis.natoque.penatibus@quismassaMauris.edu","363-6976 Lacus. St.","Quisque Fringilla PC"
7,"Merrill Norton","at.sem@elementum.net","P.O. Box 452, 6951 Egestas. St.","Nec Metus Institute"
8,"Nathaniel Carrillo","eget@massa.co.uk","Ap #438-604 Tellus St.","Blandit Viverra Corporation"
9,"Warren Valenzuela","tempus.scelerisque.lorem@ornare.co.uk","Ap #590-320 Nulla Av.","Ligula Aliquam Erat Incorporated"
10,"Donovan Hill","facilisi@augue.org","979-6729 Donec Road","Turpis In Condimentum Associates"
11,"Kamal Matthews","augue.ut@necleoMorbi.org","Ap #530-8214 Convallis, St.","Tristique Senectus Et Foundation"

7.接收Hive数据

如果上面所有的步骤都完成了，那么现在就可以从Kafka发送数据，可以看到数据流在几秒钟内就会被发送到Hive表。