Flink 集群搭建(基于flink on YARN模式)

YARN调度与资源管理知识点：yarn生产详解（资源管理+调度器）

https://blog.csdn.net/shell33168/article/details/87928001    参考该博客

基于Flink on YARN的flink集群规划(flink1.9.0与Hadoop2.8.5整合)

主机名	IP	说明
centoshadoop1	192.168.227.140	StandaloneSessionClusterEntrypoint(主节点进程名称)
centoshadoop2	192.168.227.141	TaskManagerRunner(从节点进程名称)
centoshadoop3	192.168.227.142	TaskManagerRunner(从节点进程名称)
centoshadoop4	192.168.227.143	TaskManagerRunner(从节点进程名称)

Vi ~/.bash_profile

export FLINK_HOME=/home/hadoop/flink/flink-1.9.0
export PATH=$PATH:$FLINK_HOME/bin

source ~/.bash_profile

创建flink在hadoop上的逻辑数据目录

hadoop fs  -mkdir  -p  /flink/log/hadoop-flink

解压flink-1.9.0-bin-scala_2.11.tgz  到/home/hadoop/flink目录

tar -zxvf flink-1.9.0-bin-scala_2.11.tgz -C  ~/flink

配置基本参数(结合机器情况调整)

cd /home/hadoop/flink/flink-1.9.0/conf

配置flink-conf.yaml,根据如下表格进行调整相应的参数

基础配置

参数	值	说明
jobmanager.rpc.address	centoshadoop1	jobmanager所在节点
jobmanager.rpc.port	6123	jobManager端口,默认为6123
jobmanager.heap.size	4048m	jobmanager可用内存
taskmanager.heap.size	4096m	每个TaskManager可用内存,根据集群情况指定
taskmanager.numberOfTaskSlots	3	每个taskmanager的并行度（5以内）
parallelism.default	2	启动应用的默认并行度（该应用所使用总的CPU数）
rest.port	8085	Flink web UI默认端口与spark的端口8081冲突,更改为8085

history server配置

jobmanager.archive.fs.dir	hdfs://mycluster/flink/log/hadoop-flink	因为配置了hadoop的HA,所以hdfs nameservices 指定为mycluster
historyserver.web.address	Centoshadoop1	historyserver web UI地址（需要在本地hosts文件中指定该映射关系）
historyserver.web.port	18082	historyserver web UI端口
historyserver.archive.fs.dir	hdfs://mycluster/flink/log/hadoop-flink	值与 jobmanager.archive.fs.dir 保持一致
historyserver.archive.fs.refresh-interval	10000	history server页面默认刷新时长
参数	值	说明

Flink配置文件——slaves

vi slaves

centoshadoop2

centoshadoop3

centoshadoop4

分发到其他节点

cd ~

scp -r flink/  hadoop@centoshadoop2:~/

scp -r flink/  hadoop@centoshadoop3:~/

scp -r flink/  hadoop@centoshadoop4:~/

cd /home/hadoop/flink/flink-1.9.0

bin/start-cluster.sh             -- 启动集

执行下面的命令

[hadoop@centoshadoop1 flink-1.9.0]$

cd /home/hadoop/flink/flink-1.9.0/

 bin/bin/flink run -m yarn-cluster ./examples/batch/WordCount.jar

记得监控日志: tail -f flink-hadoop-client-centoshadoop1.log 

------------------------------------------------------------

The program finished with the following exception:

java.lang.RuntimeException: Could not identify hostname and port in 'yarn-cluster'.

原因：Flink1.8以后，FIX了FLINK-11266,将flink的包中对hadoop版本进行了剔除，导致flink中直接缺少hadoop的client相关类，无法解析yarn-cluster参数。

cd /home/hadoop/flink/flink-1.9.0/bin

在config.sh配置文件 中添加HADOOP_CLASSPATH

export HADOOP_HOME=/home/hadoop/hadoop-ha/hadoop/hadoop-2.8.5
export HADOOP_CLASSPATH=$HADOOP_HOME/bin/hadoop
export HADOOP_CONF_DIR=$HADOOP_HOME/etc/hadoop

下载flink依赖的hadoop相关flink-shaded-hadoop-2-uber-2.8.3-10.0.jar包,添加到${FLINK_HOME}/lib下

https://repo.maven.apache.org/maven2/org/apache/flink/flink-shaded-hadoop-2-uber/2.8.3-10.0/flink-shaded-hadoop-2-uber-2.8.3-10.0.jar
再次执行如下命令:

./bin/flink run -m yarn-cluster ./examples/batch/WordCount.jar

报错信息：

修改hadoop安装目录下的yarn-site.xml文件

         yarn.scheduler.minimum-allocation-mb

         4048

         default value is 1024

    

    

         yarn.nodemanager.vmem-check-enabled

         false

    

     

         yarn.timeline-service.enabled

         false

    

分发该配置文件到hadoop的各个节点

 cd ~/hadoop-ha/hadoop/hadoop-2.8.5/

 scp -r yarn-site.xml  hadoop@centoshadoop2:/home/hadoop/hadoop-ha/hadoop/hadoop-2.8.5/etc/hadoop
 scp -r yarn-site.xml  hadoop@centoshadoop3:/home/hadoop/hadoop-ha/hadoop/hadoop-2.8.5/etc/hadoop
 scp -r yarn-site.xml  hadoop@centoshadoop4:/home/hadoop/hadoop-ha/hadoop/hadoop-2.8.5/etc/hadoop

重启flink集群

bin/start-cluster.sh

再次执行如下命令:

 bin/flink run -m yarn-cluster -yn 2 ./examples/batch/WordCount.jar

输出日志如下：

启动 historyServer

$ ./historyserver.sh start   启动历史资源

http://centoshadoop1:18082/

至此基于YARN的flink集群搭建完毕