前言
Apache Flink是一个开源的流处理框架,提供了分布式的、高性能的、高可用的特性,同时能够为流式应用程序提供多种编程语言的API,更多有关Flink的内容,不再累述,请参考官方文档,本文主要以Flink Standalone集群安装配置、编程实践为主。
Flink集群安装配置
首先,选择一主二从共3个节点来安装配置Flink Standalone集群:
Master:ali-bj01-tst-cluster-001.xiweiai.cnWorker:ali-bj01-tst-cluster-002.xiweiai.cnWorker:ali-bj01-tst-cluster-003.xiweiai.cn
为了方便安装文件远程拷贝,单独创建一个hadoop用户,并打通从Master节点使用ssh到Worker节点之间的免密码登录。在Master节点上下载、准备Flink 1.4.0安装文件,执行如下命令:
wget http://mirror.bit.edu.cn/apache/flink/flink-1.4.0/flink-1.4.0-bin-hadoop26-scala_2.11.tgztar xvzf flink-1.4.0-bin-hadoop26-scala_2.11.tgz
接着,修改Flink配置文件flink-1.4.0/conf/flink-conf.yaml,修改后的内容如下所示:
env.java.home: /usr/local/java/jdk1.8.0_131/jobmanager.rpc.address: ali-bj01-tst-cluster-001.xiweiai.cn# The heap size for the JobManager JVMjobmanager.heap.mb: 2048# The heap size for the TaskManager JVMtaskmanager.heap.mb: 2048
通过设置env.java.home属性,指定JAVA_HOME环境变量的值;jobmanager.rpc.address表示Flink集群的JobManager的RPC地址;设置jobmanager.heap.mb和taskmanager.heap.mb分别指定JobManager和TaskManager的堆内存大小,这个需要根据实际节点的资源情况进行配置,这里每个Worker节点都配置为2G大小。然后,修改conf/slaves文件,增加Worker节点的主机名或IP列表,内容如下所示:
ali-bj01-tst-cluster-002.xiweiai.cnali-bj01-tst-cluster-003.xiweiai.cn
最后,将配置完成的Flink安装文件,远程拷贝到各个Worker节点:
scp -r /mnt/bd/data/flink/flink-1.4.0 ali-bj01-tst-cluster-002.xiweiai.cn:/mnt/bd/data/flink/scp -r /mnt/bd/data/flink/flink-1.4.0 ali-bj01-tst-cluster-003.xiweiai.cn:/mnt/bd/data/flink/
启动集群
启动Flink集群,执行如下命令:
bin/start-cluster.sh
它会在Master节点上启动JobManager进程,然后ssh到Worker节点上启动各个TaskManager进程。当然可选地,也可以单独启动JobManager和TaskManager,在Master节点上启动JobManager:
bin/jobmanager.sh start
在各个Worker节点上启动TaskManager:
bin/taskmanager.sh start
启动完成后,可以查看对应的日志,验证Flink集群是否启动成功:
tail -100f log/flink-hadoop-jobmanager-0-ali-bj01-tst-cluster-001.xiweiai.cn.logtail -100f log/flink-hadoop-taskmanager-0-ali-bj01-tst-cluster-002.xiweiai.cn.logtail -100f log/flink-hadoop-taskmanager-0-ali-bj01-tst-cluster-003.xiweiai.cn.log
或者,查看Flink Web Dashboard UI来确认,通过浏览器打开链接:http://ali-bj01-tst-cluster-001.xiweiai.cn:8081/,如下图所示:
可见,TaskManager都已经注册到JobManager,整个Flink集群具有了计算资源(Task Slot),集群启动成功。
创建Flink项目
通过Flink官方文档,可以看到如何创建quickstart项目,假设我们使用Scala来编写Flink程序,默认使用sbt来管理依赖和程序构建,可以执行如下命令来创建:
bash
可以根据向导提示,一步一步进行(示例是我们创建的flink-demo项目):
➜ bash
基于该Demo项目,就可以开发我们自己的Flink程序了。下面,我们分别选择使用Flink提供的API来开发批量数据处理程序、流式数据处理程序、基于窗口(Windowing)的数据处理,来实现通过Flink API进行编程并提交运行。
批量数据处理
我们基于Flink发行包自带的example代码,修改WordCount程序为HdfsWordCount,支持从HDFS读取文件,然后将计算结果写入到HDFS文件,非常简单,代码如下所示:
import org.apache.flink.api.scala._ object HdfsWordCount { def main(args: Array[String]) { if(args.length != 2) { println("Parameter error!") println("Usage: org.shirdrn.flink.HdfsWordCount ") System.exit(-1) } val inputFile = args(0) val outputFile = args(1) // set up the execution environment val env = ExecutionEnvironment.getExecutionEnvironment // get input data val text = env.readTextFile(inputFile) val counts = text.flatMap { _.toLowerCase.split("W+") } .map { (_, 1) } .groupBy(0) .sum(1) // execute and save result counts .map(t => Seq(t._1, t._2).mkString("")) .writeAsText(outputFile) val jobName = getClass.getSimpleName env.execute(jobName) }}
上面代码中,env.readTextFile(inputFile)返回了DataSet[String],表示Flink中对String类型数据元素的数据集的抽象;写入HDFS,调用DataSet的writeAsText方法,传递一个HDFS路径即可。构建打包程序,执行如下命令:
sbt clean assembly
可以看到,生成了文件target/scala-2.11/flink-demo-assembly-0.0.1-SNAPSHOT.jar,可以将其提交到Flink集群上运行。有两种方式,一种方式是,通过Flink Web Dashboard,在页面上上传上述jar文件,然后配置输入的参数值,不做过多说明;另一种方式是,通过命令行的方式提交,这种方式比较方便,对应的提交运行命令,如下所示:
bin/flink run --class org.shirdrn.flink.HdfsWordCount flink-demo-assembly-0.0.1-SNAPSHOT.jar hdfs://ali-bj01-tst-cluster-001.xiweiai.cn:8020/data/beam/input/Notre-Dame-de-Paris.txt hdfs://ali-bj01-tst-cluster-001.xiweiai.cn:8020/user/hadoop/count_result.txt
输入和输出文件,都要加上HDFS前缀: hdfs://ali-bj01-tst-cluster-001.xiweiai.cn:8020,指明HDFS集群的地址。提交运行后,可以在Flink Web Dashboard看到对应的Job运行情况,可以通过命令查看输出结果文件的内容:
hdfs dfs -cat /user/hadoop/count_result.txt
流式数据处理
示例中,我们打算从一个具有3节点的Kakfa集群的topic中读取数据,数据来自网站用户浏览的Apache Server的日志,我们需要将日志行中有用的信息抽取出来,然后再写入到一个新的Kakfa topic中,供下游的其它程序处理,比如可以使用的场景有:实时推荐、实时监控,等等。首先,创建2个Kafka topic,执行如下命令行:
kafka-topics --zookeeper 172.16.117.63:2181,172.16.117.64:2181,172.16.117.65:2181 --partitions 3 --replication-factor 2 --create --topic raw_apache_log_eventkafka-topics --zookeeper 172.16.117.63:2181,172.16.117.64:2181,172.16.117.65:2181 --partitions 3 --replication-factor 2 --create --topic etld_user_behavior_event
raw_apache_log_event存储原始日志记录,etld_user_behavior_event存储经过抽取、过滤后的记录。然后,基于上面创建的flink-demo项目,在build.sbt中添加Flink Streaming及其Kakfa connector依赖,如下所示:
val flinkDependencies = Seq( "org.apache.flink" %% "flink-scala" % flinkVersion % "provided", "org.apache.flink" %% "flink-streaming-scala" % flinkVersion % "provided", "org.apache.flink" %% "flink-connector-kafka-0.9" % flinkVersion)
接着,可以继续开发代码,实现我们计划的逻辑,代码如下所示:
package org.shirdrn.flink import java.util.regex.Patternimport org.apache.flink.api.common.restartstrategy.RestartStrategiesimport org.apache.flink.api.common.serialization.SimpleStringSchemaimport org.apache.flink.api.java.utils.ParameterToolimport org.apache.flink.streaming.api.scala._import org.apache.flink.streaming.connectors.kafka.{FlinkKafkaConsumer09, FlinkKafkaProducer09} /** * Read raw events from a Kafka topic, and then write ETLed events to another Kafka topic. */object KafkaStreamingETL { private val LINE_PATTERN = Pattern.compile( "([^s]+)s-s-s[([^]]+)]s"([^"]+)"s([^s]+)s-s"([^"]+).*(d+)$" ); private def checkParams(params: ParameterTool) = { if (params.getNumberOfParameters < 5) { println("Missing parameters!" + "Usage: Kafka " + "--raw-topic " + "--etled-topic " + "--bootstrap.servers " + "--zookeeper.connect " + "--group.id ") System.exit(-1) } } def main(args: Array[String]): Unit = { // parse & check input arguments val params = ParameterTool.fromArgs(args) checkParams(params) val env = StreamExecutionEnvironment.getExecutionEnvironment configureEnv(params, env) // create a Kafka streaming source consumer for Kafka 0.9.x val kafkaConsumer = new FlinkKafkaConsumer09( params.getRequired("raw-topic"), new SimpleStringSchema, params.getProperties ) // process streaming events val messageStream = env .addSource(kafkaConsumer) .map(parseEvent(_)) .filter(!_.isEmpty) // create a Kafka producer for Kafka 0.9.x val kafkaProducer = new FlinkKafkaProducer09( params.getRequired("etled-topic"), new SimpleStringSchema, params.getProperties ) // write data into Kafka messageStream.addSink(kafkaProducer) env.execute("KafkaStreamingETL") } private def configureEnv(params: ParameterTool, env: StreamExecutionEnvironment) = { env.getConfig.disableSysoutLogging val restartAttempts = 3 val delayBetweenAttemptsMills = 30000 env.getConfig.setRestartStrategy(RestartStrategies.fixedDelayRestart( restartAttempts, delayBetweenAttemptsMills)) // create a checkpoint every 5 seconds val checkpointIntervalMillis = 1000 env.enableCheckpointing(checkpointIntervalMillis) // make parameters available in the web interface env.getConfig.setGlobalJobParameters(params) } // processed event line example: // 123.150.182.192 28/Dec/2017:01:05:21 /wp-content/plugins/akismet/_inc/form.js?ver=3.0.1 304 http://myhost.com/archives/30025.html 1 def parseEvent(msg: String): String = { val m = LINE_PATTERN.matcher(msg) if(m.find()) { val ip = m.group(1) val dt = m.group(2) val resource = m.group(3) val responseCode = m.group(4) val pageUrl = m.group(5) val timeTaken = m.group(6) Seq( ip, dt.split("s+")(0), resource.split("s+")(1), responseCode, pageUrl, timeTaken ).mkString("") } else { "" } } }
上面代码中,env.addSource(kafkaConsumer)返回了DataStream[T],后续map和filter操作都作用在该DataStream[T]上,DataStream[T]是Flink对流式数据集的抽象表示。构建打包后,提交到Flink集群上运行,执行如下命令:
bin/flink run --parallelism 3 --class org.shirdrn.flink.KafkaStreamingETL flink-demo-assembly-0.0.1-SNAPSHOT.jar --raw-topic raw_apache_log_event --etled-topic etld_user_behavior_event --zookeeper.connect 172.16.117.63:2181,172.16.117.64:2181,172.16.117.65:2181 --bootstrap.servers ali-bj01-tst-cluster-002.xiweiai.cn:9092,ali-bj01-tst-cluster-003.xiweiai.cn:9092,ali-bj01-tst-cluster-004.xiweiai.cn:9092 --group.id G_XW_WLD_ETL
我们可以在Flink Web Dashboard上查看提交的KafkaStreamingETL的运行状况,如下图所示:
也可以通过Kakfa自带的topic消费脚本工具,确认Streaming程序处理结果:
kafka-console-consumer --zookeeper 172.16.117.63:2181,172.16.117.64:2181,172.16.117.65:2181 --topic raw_apache_log_event --from-beginning
基于窗口操作(Windowing)的数据处理
我们基于前面经过ETL解析处理的日志事件记录,进行实时Windowing操作。从Kafka中读取数据,然后经过对IP进行Windowing和统计,结果输出到另一个新的Kakfa topic中,实现代码,如下所示:
object WindowedUserBrowseringAnalytics { def checkParams(params: ParameterTool) = { if (params.getNumberOfParameters < 6) { println("Missing parameters!" + "Usage: Kafka " + "--etled-topic " + "--windowed-topic " + "--bootstrap.servers " + "--zookeeper.connect " + "--group.id " + "--window-size ") System.exit(-1) } } def main(args: Array[String]): Unit = { // parse & check input arguments val params = ParameterTool.fromArgs(args) checkParams(params) val env = StreamExecutionEnvironment.getExecutionEnvironment configureEnv(params, env) val windowSize = params.getRequired("window-size").toLong val kafkaConsumer = new FlinkKafkaConsumer09( params.getRequired("etled-topic"), new SimpleStringSchema, params.getProperties ) val kafkaProducer = new FlinkKafkaProducer09( params.getRequired("windowed-topic"), new SimpleStringSchema, params.getProperties ) val stream: DataStream[String] = env.addSource(kafkaConsumer) stream .map(event => (event.split("")(0), 1L)) .keyBy(0) .window(TumblingEventTimeWindows.of(Time.seconds(windowSize))) .reduce((event1, event2) => (event1._1, event1._2 + event2._2)) .map(r => { val df = new SimpleDateFormat("yyyy-MM-dd:HH:mm") Seq(df.format(new Date()), r._1, r._2).mkString("") }) .addSink(kafkaProducer) env.execute(getClass.getSimpleName) }}
构建打包,提交到Flink集群运行:
bin/flink run --parallelism 2 --class org.shirdrn.flink.WindowedUserBrowseringAnalytics flink-demo-assembly-0.0.1-SNAPSHOT.jar --etled-topic etld_user_behavior_event --windowed-topic windowed_analytics_result --zookeeper.connect 172.16.117.63:2181,172.16.117.64:2181,172.16.117.65:2181 --bootstrap.servers ali-bj01-tst-cluster-002.xiweiai.cn:9092,ali-bj01-tst-cluster-003.xiweiai.cn:9092,ali-bj01-tst-cluster-004.xiweiai.cn:9092 --window-size 60 --group.id G_XW_MON_ALTS
可以通过前面类似的方式,查看Job运行状况和结果。(时延军(包含链接:http://shiyanjun.cn))
可以转发关注小编,每天更新技术好文!
感谢大家支持!!