Dolphin Scheduler 1.2.0 部署参数分析

Dolphin Scheduler目录配置文件解读

(讲解配置文件的作用,具体配置在install.sh部署文件中完成)
image.png

  • bin 启动脚本
  • conf 配置文件
  • lib ds依赖的jar包
  • script 数据库创建升级脚本,部署分发脚本
  • sql ds的元数据创建升级sql文件
  • install脚本 部署ds主要的配置文件修改处

bin

bin目录下比较重要的是dolphinscheduler-daemon文件,之前版本中极容易出现的找不到jdk问题来源,当前版本的jdk已经export了本机的$JAVA_HOME,再也不用担心找不到jdk了。
image.png

conf

非常重要的配置文件目录!!!
非常重要的配置文件目录!!!
非常重要的配置文件目录!!!
image.png

  • env目录下的.dolphinscheduller_env.sh文件中记录了所有跟ds-task相关的环境变量,1.2.0版本的Spark不具备指定Spark版本的功能,可以注释掉SPARK_HOME1或者将SPARK_HOME1和SPARK_HOME2均配置为集群中的Spark2。下面给出CDH中的配置,测试环境中没有部署Flink,请忽略Flink的配置。(特别注意这是个隐藏文件,需要ls -al)
export HADOOP_HOME=/opt/cloudera/parcels/CDH/lib/hadoop
export HADOOP_CONF_DIR=/opt/cloudera/parcels/CDH/lib/hadoop/etc/hadoop
#可以注释掉,也可以配置为SPARK_HOME2
#export SPARK_HOME1=/opt/cloudera/parcels/SPARK2/lib/spark2
export SPARK_HOME2=/opt/cloudera/parcels/SPARK2/lib/spark2
export PYTHON_HOME=/usr/local/anaconda3/bin/python
export JAVA_HOME=/usr/java/jdk1.8.0_131
export HIVE_HOME=/opt/cloudera/parcels/CDH/lib/hive
export FLINK_HOME=/opt/soft/flink
export PATH=$HADOOP_HOME/bin:$SPARK_HOME1/bin:$SPARK_HOME2/bin:$PYTHON_HOME:$JAVA_HOME/bin:$HIVE_HOME/bin:$PATH:$FLINK_HOME/bin:$PATH

common目录

common目录包含:common.properties和hadoop/hadoop.properties

  • common.properies
    • ds的task队列实现方式,默认是Zookeeper
    • ds的task和资源的worker执行路径
    • 资源中心
      • 资源中心可选择HDFS和S3
    • 资源文件类型
    • kerberos
    • 开发状态
      • 开发测试可以开启,生产环境建议设置为false
    • ds的环境变量配置,本地调试的时候,需要保证dolphinscheduler.env.path存在
  • hadoop.properties
    • hdfs namenode配置
      • 单点可以直接写namenode的ip
      • hdfsHA需要将集群的core-site.xml和hdfs-site.xml文件拷贝到ds的conf目录下
    • s3配置
    • yarn resourcemanager配置
      • 单点配置yarn.application.status.address
      • HA配置yarn.resourcemanager.ha.rm.ids

config目录

config目录包含install_config.conf和run_config.conf

  • install_config.conf
    • ds的安装路径
    • 部署用户
    • 部署ds的机器组ip
  • run_config.conf
    • 指定ds的masters,workers,alertServer,apiServer部署在哪些机器上

alert.properties

  • 邮件告警配置
  • excel下载目录
  • 企业微信配置

application-api.properties

  • apiserver端口,上下文,日志等

application-dao.properties

敲黑板,重点!!!ds的元数据库配置,在ds-1.2.0中默认的数据库是pg,如果要使用MySQL,需要将MySQL的jdbc包放到lib目录下。

  • ds元数据库配置

master.properties

  • master执行线程数
  • master并行任务上限
  • master资源CPU和内存阈值,超出阈值不会进行dag切分

worker.properties

  • worker执行线程数
  • worker一次提交任务数
  • worker资源CPU和内存阈值,超出不会去task队列拉取task

Zookeeper.properties

  • zk集群
  • ds所需zk的znode,包含dag和task的分布式锁和master和worker的容错

quartz.properties

ds的定时由quartz框架完成,特别注意里边有quartz的数据库配置!!!

  • quartz的基本属性,线程池和job配置
  • quartz元数据库配置

install脚本

install.sh部署脚本是ds部署中的重头戏,下面将参数分组进行分析。

数据库配置

# for example postgresql or mysql ...
dbtype="postgresql"

# db config
# db address and port
dbhost="192.168.xx.xx:5432"

# db name
dbname="dolphinscheduler"

# db username
username="xx"

# db passwprd
# Note: if there are special characters, please use the \ transfer character to transfer
passowrd="xx"
  • dbtype参数可以设置postgresql和mysql,这里指定了ds连接元数据库的jdbc相关信息

部署用户&目录

# conf/config/install_config.conf config
# Note: the installation path is not the same as the current path (pwd)
installPath="/data1_1T/dolphinscheduler"

# deployment user
# Note: the deployment user needs to have sudo privileges and permissions to operate hdfs. If hdfs is enabled, the root directory needs to be created by itself
deployUser="dolphinscheduler"
  • installPath是安装路径,在执行install.sh之后,会把ds安装到指定目录,如/opt/ds-agent。installPath不要和当前要一键安装的install.sh是同一目录。
  • deployUser是指ds的部署用户,该用户需要在部署ds的机器上打通sudo免密,并且需要具有操作hdfs的权限,建议挂到hadoop的supergroup组下。

zk集群&角色指定

  • 配置zk集群的时候,特别注意:要用ip:2181的方式配置上去,一定要把端口带上。
  • ds一共包括master worker alert api四种角色,其中alert api只需指定一台机器即可,master和worker可以部署多态机器。下面的例子就是在4台机器中,部署2台master,2台worker,1台alert,1台api
  • ips参数,填写所有需要部署机器的hostname
  • masters,填写部署master机器的hostname
  • workers,填写部署worker机器的hostname
  • alertServer,填写部署alert机器的hostname
  • apiServers,填写部署api机器的hostname
  • zkroot参数可以通过调整,在一套zk集群中,托管多个ds集群,如配置zkRoot="/dspro",zkRoot="/dstest"
# zk cluster
zkQuorum="192.168.xx.xx:2181,192.168.xx.xx:2181,192.168.xx.xx:2181"

# install hosts
# Note: install the scheduled hostname list. If it is pseudo-distributed, just write a pseudo-distributed hostname
ips="ark0,ark1,ark2,ark3"

# conf/config/run_config.conf config
# run master machine
# Note: list of hosts hostname for deploying master
masters="ark0,ark1"

# run worker machine
# note: list of machine hostnames for deploying workers
workers="ark2,ark3"

# run alert machine
# note: list of machine hostnames for deploying alert server
alertServer="ark3"

# run api machine
# note: list of machine hostnames for deploying api server
apiServers="ark1"

# zk config
# zk root directory
zkRoot="/dolphinscheduler"

# used to record the zk directory of the hanging machine
zkDeadServers="$zkRoot/dead-servers"

# masters directory
zkMasters="$zkRoot/masters"

# workers directory
zkWorkers="$zkRoot/workers"

# zk master distributed lock
mastersLock="$zkRoot/lock/masters"

# zk worker distributed lock
workersLock="$zkRoot/lock/workers"

# zk master fault-tolerant distributed lock
mastersFailover="$zkRoot/lock/failover/masters"

# zk worker fault-tolerant distributed lock
workersFailover="$zkRoot/lock/failover/workers"

# zk master start fault tolerant distributed lock
mastersStartupFailover="$zkRoot/lock/failover/startup-masters"

# zk session timeout
zkSessionTimeout="300"

# zk connection timeout
zkConnectionTimeout="300"

# zk retry interval
zkRetrySleep="100"

# zk retry maximum number of times
zkRetryMaxtime="5"

邮件配置&excel文件路径

  • 邮件配置这块也是大家非常容易出问题的,建议可以拉一下ds的代码,跑一下alert.MailUtilisTest这个测试类,下面给出QQ邮箱配置方式。如果是内网邮箱,需要注意的是ssl是否需要关闭,以及mail.user登陆用户是否需要去掉邮箱后缀。
  • excel路径则需要保证该路径的写入权限
#QQ邮箱配置
# alert config
# mail protocol
mailProtocol="SMTP"

# mail server host
mailServerHost="smtp.qq.com"

# mail server port
mailServerPort="465"

# sender
mailSender="783xx8369@qq.com"

# user
mailUser="783xx8369@qq.com"

# sender password
mailPassword="邮箱授权码"

# TLS mail protocol support
starttlsEnable="false"

sslTrust="smtp.qq.com"

# SSL mail protocol support
# note: The SSL protocol is enabled by default. 
# only one of TLS and SSL can be in the true state.
sslEnable="true"

# download excel path
xlsFilePath="/tmp/xls"

# alert port
alertPort=7789

apiServer配置

  • apiServer这里可以关注一下,apiserver的端口和上下文即apiServerPort和apiServerContextPath参数
# api config
# api server port
apiServerPort="12345"

# api session timeout
apiServerSessionTimeout="7200"

# api server context path
apiServerContextPath="/dolphinscheduler/"

# spring max file size
springMaxFileSize="1024MB"

# spring max request size
springMaxRequestSize="1024MB"

# api max http post size
apiMaxHttpPostSize="5000000"

资源中心&YARN

  • ds的资源中心支持HDFS和S3.
  • resUploadStartupType="HDFS"则开启hdfs作为资源中心。
  • defaultFS,如果hdfs没有配置HA则需要在这里写上单点namenode的ip,如果HDFS是HA则需要将集群的core-site.xml文件和hdfs-site.xml文件拷贝到conf目录下
  • yarnHaIps,如果yarn启用了HA,配置两个resourcemanager的ip,如果是单点,配置空字符串
  • singleYarnIp,如果yarn是单点,配置resourcemanager的ip
  • hdfsPath,HDFS上ds存储资源的根路径,可采用默认值,如果是从1.1.0版本进行升级,需要注意这个地方,改为/escheduler
# resource Center upload and select storage method:HDFS,S3,NONE
resUploadStartupType="NONE"

# if resUploadStartupType is HDFS,defaultFS write namenode address,HA you need to put core-site.xml and hdfs-site.xml in the conf directory.
# if S3,write S3 address,HA,for example :s3a://dolphinscheduler,
# Note,s3 be sure to create the root directory /dolphinscheduler
defaultFS="hdfs://mycluster:8020"

# if S3 is configured, the following configuration is required.
s3Endpoint="http://192.168.xx.xx:9010"
s3AccessKey="xxxxxxxxxx"
s3SecretKey="xxxxxxxxxx"

# resourcemanager HA configuration, if it is a single resourcemanager, here is yarnHaIps=""
yarnHaIps="192.168.xx.xx,192.168.xx.xx"

# if it is a single resourcemanager, you only need to configure one host name. If it is resourcemanager HA, the default configuration is fine.
singleYarnIp="ark1"

# hdfs root path, the owner of the root path must be the deployment user. 
# versions prior to 1.1.0 do not automatically create the hdfs root directory, you need to create it yourself.
hdfsPath="/dolphinscheduler"

# have users who create directory permissions under hdfs root path /
# Note: if kerberos is enabled, hdfsRootUser="" can be used directly.
hdfsRootUser="hdfs"

开发状态

  • devState在测试环境部署的时候可以调为true,生产环境部署建议调为false
# development status, if true, for the SHELL script, you can view the encapsulated SHELL script in the execPath directory. 
# If it is false, execute the direct delete
devState="true"

角色参数

  • 下面的参数主要是调整的application.properties里边的配置,涉及master,worker和apiserver
  • apiServerPort可以自定义修改apiserver的端口,注意需要跟前端保持一致。
  • master和worker的参数,初次部署建议保持默认值,如果在运行当中出现性能问题在作调整,有条件可以压一下自身环境中的master和worker的最佳线程数。
  • worker.reserved.memory是worker的内存阈值,masterReservedMemory是master的内存阈值,建议调整为0.1
  • masterMaxCpuLoadAvg建议注释掉,ds-1.2.0master和worker的CPU负载给出了默认cpu线程数 * 2的默认值
# master config 
# master execution thread maximum number, maximum parallelism of process instance
masterExecThreads="100"

# the maximum number of master task execution threads, the maximum degree of parallelism for each process instance
masterExecTaskNum="20"

# master heartbeat interval
masterHeartbeatInterval="10"

# master task submission retries
masterTaskCommitRetryTimes="5"

# master task submission retry interval
masterTaskCommitInterval="100"

# master maximum cpu average load, used to determine whether the master has execution capability
#masterMaxCpuLoadAvg="10"

# master reserve memory to determine if the master has execution capability
masterReservedMemory="1"

# master port
masterPort=5566

# worker config 
# worker execution thread
workerExecThreads="100"

# worker heartbeat interval
workerHeartbeatInterval="10"

# worker number of fetch tasks
workerFetchTaskNum="3"

# worker reserve memory to determine if the master has execution capability
workerReservedMemory="1"

# master port
workerPort=7788

特别注意

  • ds需要启用资源中心之后,才可以创建租户,因此资源中心的配置一定要正确
  • ds老版本部署需要配置JDK的问题已经解决
  • installPath不要和当前要一键安装的install.sh是同一目录
  • ds的task运行都依赖env目录下的环境变量文件,需要正确配置
  • HDFS高可用,需要把core-site.xml和hdfs-site.xml文件拷贝到conf目录下
  • 邮件配置中mailUser和mailSender的区别