Gobblin采集kafka数据

一.Gobblin环境变量准备

需要配置好Gobblin0.11.0工作时对应的环境变量，可以去Gobblin的bin目录的gobblin-env.sh配置，比如

job.name=GobblinKafkaQuickStart
job.group=GobblinKafka
job.description=Gobblin quick start job for Kafka
job.lock.enabled=false
job.schedule=0/3 * * * ?
kafka.brokers=101.236.39.141:9092,101.236.46.114:9092,101.236.46.113:9092
source.class=gobblin.source.extractor.extract.kafka.KafkaSimpleSource
extract.namespace=gobblin.extract.kafka
     
writer.builder.class=gobblin.writer.SimpleDataWriterBuilder
writer.file.path.type=tablename
writer.destination.type=HDFS
writer.output.format=txt
     
data.publisher.type=gobblin.publisher.BaseDataPublisher
     
mr.job.max.mappers=1
     
metrics.reporting.file.enabled=true
metrics.log.dir=${env:GOBBLIN_WORK_DIR}/metrics
metrics.reporting.file.suffix=txt
     
bootstrap.with.offset=earliest

这里需要配置好抽取数据的kafka broker以及一些gobblin的工作组件，如source,extract,writer,publisher等，不明白的可以参考Gobblin wiki,很详细.

我这里额外配置了一个job.schedule让gobblin三分钟检查一次kafka的所有topic是否有新增，然后抽取任务就会三分钟一次定时执行.这里用的Gobblin自带的Quartz定时器.

ok,配置好以后进入Gobblin根目录,启动命令如:

 bin/gobblin-standalone.sh –conffile $GOBBLIN_JOB_CONFIG_DIR/gobblinStandalone.pull start

我这里GOBBLIN_JOB_CONFIG_DIR有多个pull文件，因此需要指明，如果GOBBLIN_JOB_CONFIG_DIR下只有一个配置文件，那么直接bin/gobblin-standalone.sh start即可执行

最终抽取过来的数据会输出到GOBBLIN_WORK_DIR/job-output 中去.

三.Gobblin MapReduce模式配置和使用

这次配置Gobblin会使用MapReduce来抽取kafka数据到Hdfs，新建gobblin-mr.pull文件，配置如下

job.name=GobblinKafkaMapreduce
job.group=GobblinKafkaForMapreduce
job.description=Gobblin quick start job for Kafka
job.lock.enabled=false
kafka.brokers=101.236.39.141:9092,101.236.46.114:9092,101.236.46.113:9092

source.class=gobblin.source.extractor.extract.kafka.KafkaSimpleSource
extract.namespace=gobblin.extract.kafka
topic.whitelist=boot

writer.builder.class=gobblin.writer.SimpleDataWriterBuilder
writer.file.path.type=tablename
writer.destination.type=HDFS
writer.output.format=txt
simple.writer.delimiter=

data.publisher.type=gobblin.publisher.BaseDataPublisher

mr.job.max.mappers=1

metrics.reporting.file.enabled=true
metrics.log.dir=${env:GOBBLIN_WORK_DIR}/metrics
metrics.reporting.file.suffix=txt

bootstrap.with.offset=earliest

fs.uri=hdfs://101.236.39.141:9000
#fs.uri=hdfs://clusterYL
writer.fs.uri=${fs.uri}
state.store.fs.uri=${fs.uri}

mr.job.root.dir=/gobblin/working
state.store.dir=/gobblin/state-store
task.data.root.dir=/gobblin/task-data
data.publisher.final.dir=/gobblin/job-output

注意标红部分的配置第一行，我这里加了topic过滤，只对topic名称为jsonTest的主题感兴趣

因为需求是需要将gobblin的topic数据按照每天每小时来进行目录分区，具体分区目录需要根据kafka record中的时间字段来

我这里record是json格式的，时间字段格式如{…"time":"2016-10-12 00:30:20"…},因此需要继承Gobblin的TimeBasedWriterPartitioner来重写子类方法按照时间字段对hdfs的目录分区

以下配置需要注意

fs.uri=master:8020

改成自己的集群的hdfs地址

重写的hdfs按照json时间字段分区的子类，代码我提交到github了，参考如下链接

https://github.com/cssdongl/gobblin/blob/master/gobblin-example/src/main/java/gobblin/example/simplejson/TimeBasedJsonWriterPartitioner.java

将扩展后的类加入Gobblin相应的模块，我这里是放入gobblin-example模块中去了，重新build,build有问题的话请参考这篇文章

上面配置文件最后的那些路径都是hdfs路径，请确保Gobblin有读写权限

随后启动命令

bin/gobblin-mapreduce.sh --conf $GOBBLIN_JOB_CONFIG_DIR/gobblin-mr.pull

运行成功后，hdfs会出现如下目录,jsonTest是按照对应topic名称生成的,如下图

注意MR模式配置Quartz定时调度我试了好几次不起作用，因此如果需要定时执行抽取的话请利用外部的工具，比如Linux的crontab或者Oozie或者Azkaban都是可以的.

四.Gobblin使用总结

1>先熟悉Gobblin官方wiki,写的很详细

2>github上fork一个源代码仔细阅读下source,extract，partioner这块儿的代码

3>使用中遇到问题多研究Gobblin的log和Hadoop的log.

参考资料:

http://gobblin.readthedocs.io/en/latest/case-studies/Kafka-HDFS-Ingestion/

http://gobblin.readthedocs.io/en/latest/user-guide/Partitioned-Writers/

http://gobblin.readthedocs.io/en/latest/developer-guide/IDE-setup/

http://gobblin.readthedocs.io/en/latest/user-guide/FAQs/