Flink任务、Spark任务提交到集群,通常需要将可执行Jar上传到集群,手动执行任务提交指令,如果有配套的大数据平台则需要上传Jar,由调度系统进行任务提交。对开发者来说,本地IDEA调试Flink、Spark任务不涉及对象的序列化及反序列化,任务在本地调试通过后,执行在分布式环境下也可能会出错。而将任务提交到集群进行调试还要走那些繁琐的流程太影响效率了。
因此,为方便大数据开发人员进行快速开发调试,开发了从本地IDEA提交Flink/Spark任务到集群的工具类 flink-spark-submiter。任务提交代码稍加改造后也可以和上层调度系统进行集成,替代脚本模式进行任务提交的方式。
支持Flink yarnPerJob、Standalone 、yarnSession模式下的任务提交。
支持Spark任务以Yarn Cluster模式提交到YARN,支持自动上传用户Jar包,依赖的Spark Jars需要提前上传到HDFS。
支持Spark任务提交到K8s Cluster,执行的jar需要包含在镜像中,任务执行时需要传递镜像名称及可执行文件路径。如果需要操作hive表,则需要传递集群所在文件夹,以及HADOOP_USER_NAME,系统进行Hadoop文件的挂载及环境变量的设置。
Flink 多执行模式任务提交
- 需要填写Flink任务运行时参数配置,任务运行所在的集群配置路径,本地Flink根路径。项目依赖flink1.10版本。
- 支持以YarnSession、YarnPerjob、Standalone模式进行任务提交,返回ApplicationId。
- example模块下包含一个FlinkDemo,打包后会转移到项目的examplJars中,可以尝试进行任务提交。
- 任务提交后,根据ApplicationId获取任务执行使用的jm、tm日志基本信息,包含日志访问URL,日志总字节大小,根据日志基本信息可以做日志滚动展示,防止Yarn日志过大导致日志读取卡死。
任务提交示例:
// 可执行jar包路径
String runJarPath = "/Users/maqi/code/ClustersSubmiter/exampleJars/flink-kafka-reader/flink-kafka-reader.jar";
// 任务参数
String[] execArgs = new String[]{"-jobName", "flink110Submit", "--topic", "mqTest01", "--bootstrapServers", "172.16.8.107:9092"};
// 任务名称
String jobName = "Flink perjob submit";
// flink 文件夹路径
String flinkConfDir = "/Users/maqi/tmp/flink/flink-1.10.0/conf";
// flink lib包路径
String flinkJarPath = "/Users/maqi/tmp/flink/flink-1.10.0/lib";
// yarn 文件夹路径
// String yarnConfDir = "/Users/maqi/tmp/hadoopconf";
String yarnConfDir = "/Users/maqi/tmp/hadoopconf";
// 作业依赖的外部文件,例如:udf jar , keytab
String[] dependFile = new String[]{"/Users/maqi/tmp/flink/flink-1.10.0/README.txt"};
// 任务提交队列
String queue = "root.users.hdfs";
// flink任务执行模式
String execMode = "yarnPerjob";
// yarnsession appid配置
Properties yarnSessionConfProperties = null;
// savepoint 及并行度相关
Properties confProperties = new Properties();
confProperties.setProperty("parallelism", "1");
JobParamsInfo jobParamsInfo = JobParamsInfo.builder()
.setExecArgs(execArgs)
.setName(jobName)
.setRunJarPath(runJarPath)
.setDependFile(dependFile)
.setFlinkConfDir(flinkConfDir)
.setYarnConfDir(yarnConfDir)
.setConfProperties(confProperties)
.setYarnSessionConfProperties(yarnSessionConfProperties)
.setFlinkJarPath(flinkJarPath)
.setQueue(queue)
.build();
String applicationId = runFlinkJob(jobParamsInfo, execMode);
//任务启动后,拉取jm,tm日志相关信息。
Thread.sleep(20000);
List<String> logsInfo = new RunningLog().getRollingLogBaseInfo(jobParamsInfo, applicationId);
logsInfo.forEach(System.out::println);
jobmanager日志格式:
{
"logs":[
{
"name":"jobmanager.err ",
"totalBytes":"555",
"url":"http://172-16-10-204:8042/node/containerlogs/container_e185_1593317332045_2246_01_000002/admin/jobmanager.err/"
},
{
"name":"jobmanager.log ",
"totalBytes":"31944",
"url":"http://172-16-10-204:8042/node/containerlogs/container_e185_1593317332045_2246_01_000002/admin/jobmanager.log/"
},
{
"name":"jobmanager.out ",
"totalBytes":"0",
"url":"http://172-16-10-204:8042/node/containerlogs/container_e185_1593317332045_2246_01_000002/admin/jobmanager.out/"
}
],
"typeName":"jobmanager"
}
taskmanager日志格式:
{
"logs":[
{
"name":"taskmanager.err ",
"totalBytes":"560",
"url":"http://node03:8042/node/containerlogs/container_e27_1593571725037_0170_01_000002/admin/taskmanager.err/"
},
{
"name":"taskmanager.log ",
"totalBytes":"35937",
"url":"http://node03:8042/node/containerlogs/container_e27_1593571725037_0170_01_000002/admin/taskmanager.log/"
},
{
"name":"taskmanager.out ",
"totalBytes":"0",
"url":"http://node03:8042/node/containerlogs/container_e27_1593571725037_0170_01_000002/admin/taskmanager.out/"
}
],
"otherInfo":"{"dataPort":36218,"freeSlots":0,"hardware":{"cpuCores":4,"freeMemory":241172480,"managedMemory":308700779,"physicalMemory":8201641984},"id":"container_e27_1593571725037_0170_01_000002","path":"akka.tcp://flink@node03:36791/user/taskmanager_0","slotsNumber":1,"timeSinceLastHeartbeat":1593659561129}",
"typeName":"taskmanager"
}
Spark on yarn 任务提交
- 填写用户程序包路径、执行参数、集群配置文件夹、安全认证等相关配置。
- Spark任务提交使用Yarn cluster模式,使用的Spark Jar需要提前上传到HDFS并作为archive的参数。
- 针对SparkSQL任务,通过提交examples中的spark-sql-proxy程序包来直接操作hive表。
提交示例:
public static void main(String[] args) throws Exception {
boolean openKerberos = true;
String appName = "todd spark submit";
String runJarPath = "/Users/maqi/code/ClustersSubmiter/exampleJars/spark-sql-proxy/spark-sql-proxy.jar";
String mainClass = "cn.todd.spark.SparksqlProxy";
String yarnConfDir = "/Users/maqi/tmp/hadoopconf";
String principal = "hdfs/node1@DTSTACK.COM";
String keyTab = "/Users/maqi/tmp/hadoopconf/hdfs.keytab";
String jarHdfsDir = "sparkproxy2";
String archive = "hdfs://nameservice1/sparkjars/jars";
String queue = "root.users.hdfs";
String execArgs = getExampleJobParams();
Properties confProperties = new Properties();
confProperties.setProperty("spark.executor.cores","2");
JobParamsInfo jobParamsInfo = JobParamsInfo.builder()
.setAppName(appName)
.setRunJarPath(runJarPath)
.setMainClass(mainClass)
.setYarnConfDir(yarnConfDir)
.setPrincipal(principal)
.setKeytab(keyTab)
.setJarHdfsDir(jarHdfsDir)
.setArchivePath(archive)
.setQueue(queue)
.setExecArgs(execArgs)
.setConfProperties(confProperties)
.setOpenKerberos(BooleanUtils.toString(openKerberos, "true", "false"))
.build();
YarnConfiguration yarnConf = YarnConfLoaderUtil.getYarnConf(yarnConfDir);
String applicationId = "";
if (BooleanUtils.toBoolean(openKerberos)) {
UserGroupInformation.setConfiguration(yarnConf);
UserGroupInformation userGroupInformation = UserGroupInformation.loginUserFromKeytabAndReturnUGI(principal, keyTab);
applicationId = userGroupInformation.doAs((PrivilegedExceptionAction<String>) () -> LauncherMain.run(jobParamsInfo, yarnConf));
} else {
LauncherMain.run(jobParamsInfo, yarnConf);
}
System.out.println(applicationId);
}
Spark on k8s 任务提交
- 基于Spark2.4.4进行开发,通过将spark-sql-proxy.jar包打入镜像来执行Sparksql并操作Hive表,无其他特殊操作。
- 操作Hive时需要传递hadoopConfDir,程序会自动将.xml文件内容进行挂载,如果非root用户操作Hive,需要设置HADOOP_USER_NAME。
- 通过读取kubeConfig配置文件进行Kuberclient的创建,而非官方提供的master url方式。
- 任务提交后立即返回spark-app-selector id,从而进行POD状态获取。
public static void main(String[] args) throws Exception {
String appName = "todd spark submit";
// 镜像内的jar路径
String runJarPath = "local:///opt/dtstack/spark/spark-sql-proxy.jar";
String mainClass = "cn.todd.spark.SparksqlProxy";
String hadoopConfDir = "/Users/maqi/tmp/hadoopconf/";
String kubeConfig = "/Users/maqi/tmp/conf/k8s.config";
String imageName = "mqspark:2.4.4";
String execArgs = getExampleJobParams();
Properties confProperties = new Properties();
confProperties.setProperty("spark.executor.instances", "2");
confProperties.setProperty("spark.kubernetes.namespace", "default");
confProperties.setProperty("spark.kubernetes.authenticate.driver.serviceAccountName", "spark");
confProperties.setProperty("spark.kubernetes.container.image.pullPolicy", "IfNotPresent");
JobParamsInfo jobParamsInfo = JobParamsInfo.builder()
.setAppName(appName)
.setRunJarPath(runJarPath)
.setMainClass(mainClass)
.setExecArgs(execArgs)
.setConfProperties(confProperties)
.setHadoopConfDir(hadoopConfDir)
.setKubeConfig(kubeConfig)
.setImageName(imageName)
.build();
String id = run(jobParamsInfo);
System.out.println(id);
}