Spark spark-submit 提交的几种模式

local 模式

package com.imooc.spark.Test

import org.apache.spark.sql.types.{StringType, StructField, StructType}
import org.apache.spark.sql.{Row, SaveMode, SparkSession}

/**
  * 测试sparkContext 案例
  */
object TestOfSparkContext2 {

  def main(args: Array[String]): Unit = {
    //System.setProperty("hadoop.home.dir", "E:\\soft\\winutils\\hadoop-common-2.2.0-bin-master")
    val spark = SparkSession.builder()
      .appName("TestOfSparkContext2")
      .master("local[2]").getOrCreate()

    //反射的方式 RDD=>DF
    // reflection(spark)
    //编程的方式 RDD=>DF
    program(spark)
    spark.stop()
  }

  private def reflection(spark: SparkSession) = {
    val rdd = spark.sparkContext
      .textFile("file///home/hadoop/data/test-data-spark/emp-forSpark.txt")
      .map(w => w.split("\t"))

    import spark.implicits._
    val empDF = rdd.map(line =>
      EMP(line(0), line(1), line(2), line(3),
        line(4), line(5), line(6), line(7))).toDF()
    empDF.printSchema()
    empDF.show()
  }

  def program(spark: SparkSession) = {
    val infoRdd = spark.sparkContext
      .textFile("/home/hadoop/data/test-data-spark/emp-forSpark.txt")
      .map(w => w.split("\t")).map(line => Row(line(0), line(1), line(2), line(3), line(4), line(5), line(6), line(7)))

    val filedName = Array(StructField("empNo", StringType, true)
      , StructField("ename", StringType, true), StructField("job", StringType, true)
      , StructField("mgr", StringType, true), StructField("hireDate", StringType, true)
      , StructField("sal", StringType, true), StructField("comm", StringType, true)
      , StructField("deptNo", StringType, true))

    val schema = StructType(filedName)
    val empDf = spark.createDataFrame(infoRdd, schema)
    empDf.printSchema()
     empDf.show(false)

    //注册临时表
    empDf.createOrReplaceTempView("emp")
    val sqlDF = spark.sql("select ename,job,comm from emp")
   // sqlDF.printSchema()
    //sqlDF.show()
    /*empDf.write.format("json")
      .mode(SaveMode.Overwrite).save("E:///testData/empTable3")*/
    empDf.coalesce(1).write.format("json").mode(SaveMode.Overwrite)
      .partitionBy("deptno").save("/home/hadoop/app/spark-2.1.1-bin-2.6.0-cdh5.7.0/testData/emp-jsonTable")
  }

  case class EMP(empNo: String, empName: String, job: String, mgr: String, hiredate: String,
                 sal: String, comm: String, deptNo: String)

}

打包

提交
在$SPARK_HOME 下执行

./bin/spark-submit \
--class com.imooc.spark.Test.TestOfSparkContext2 \
--master local[2] \
/home/hadoop/data/test-jar/sql-1.0.jar

参数解释
对empDf.write.format(“json”)
.mode(SaveMode.Overwrite).save(“E:///testData/empTable3”)*/
empDf.coalesce(1).write.format(“json”).mode(SaveMode.Overwrite)
.partitionBy(“deptno”).save("/home/hadoop/app/spark-2.1.1-bin-2.6.0-cdh5.7.0/testData/emp-jsonTable")中的参数解释

write.format(“json”)    写入文件的格式
mode(SaveMode.Overwrite).    保存文件,如果文件已经存在,则覆盖.还有:Append(如果存在,则追加),ErrorIfExists(如果存在报错,则退出)等)
empDf.coalesce(1)    分区的个数,如10,则在每个部门分区下被分成10个新分区
partitionBy(“deptno”).    指定存储分区字段
save("/home/hadoop/app/spark-2.1.1-bin-2.6.0-cdh5.7.0/testData/emp-jsonTable")    报错文件路径.注意:如果是直接保存到本地,file://path…;如果是保存到HDFS下,则路径是:hdfs:/ip:port+path;特别是如果是yarn模式提交,文件路径,直接写hdfs上的路径,且只能是hdfs上的路,如:/user/hadoop/…
yarn client模式提交

package com.imooc.spark.Test

import org.apache.spark.sql.types.{StringType, StructField, StructType}
import org.apache.spark.sql.{Row, SaveMode, SparkSession}

/**
  * 测试sparkContext 案例
  */
object TestOfSparkContext2OnYarn {

  def main(args: Array[String]): Unit = {
    //System.setProperty("hadoop.home.dir", "E:\\soft\\winutils\\hadoop-common-2.2.0-bin-master")
    /*val spark = SparkSession.builder()
      .appName("TestOfSparkContext2OnYarn")
      .master("local[2]").getOrCreate()*/
    val path = "/user/hadoop/data/test-data-forSpark/emp-forSpark.txt"
    val spark = SparkSession.builder().getOrCreate()

    //反射的方式 RDD=>DF
    // reflection(spark,path)
    //编程的方式 RDD=>DF
    program(spark,path)
    spark.stop()
  }

  private def reflection(spark: SparkSession,path:String ) = {
    val rdd = spark.sparkContext
      .textFile(path)
      .map(w => w.split("\t"))

    import spark.implicits._
    val empDF = rdd.map(line =>
      EMP(line(0), line(1), line(2), line(3),
        line(4), line(5), line(6), line(7))).toDF()
    empDF.printSchema()
    empDF.show()
  }

  def program(spark: SparkSession,path:String) = {
    val infoRdd = spark.sparkContext
      .textFile(path)
      .map(w => w.split("\t")).map(line => Row(line(0), line(1), line(2), line(3), line(4), line(5), line(6), line(7)))

    val filedName = Array(StructField("empNo", StringType, true)
      , StructField("ename", StringType, true), StructField("job", StringType, true)
      , StructField("mgr", StringType, true), StructField("hireDate", StringType, true)
      , StructField("sal", StringType, true), StructField("comm", StringType, true)
      , StructField("deptNo", StringType, true))

    val schema = StructType(filedName)
    val empDf = spark.createDataFrame(infoRdd, schema)
    empDf.printSchema()
     empDf.show(false)

    //注册临时表
    empDf.createOrReplaceTempView("emp")
    val sqlDF = spark.sql("select ename,job,comm from emp")
   // sqlDF.printSchema()
    //sqlDF.show()
    /*empDf.write.format("json")
      .mode(SaveMode.Overwrite).save("E:///testData/empTable3")*/
    empDf.coalesce(1).write.format("json").mode(SaveMode.Overwrite)
      .partitionBy("deptno").save("hdfs://hadoop001:9000/user/hadoop/emp-spark-test/emp-jsonOnYarnTable")
  }

  case class EMP(empNo: String, empName: String, job: String, mgr: String, hiredate: String,
                 sal: String, comm: String, deptNo: String)

}

提交
./bin/spark-submit \
--class com.imooc.spark.Test.TestOfSparkContext2OnYarn \
--master yarn \
/home/hadoop/data/jar-test/sql-3.0-onYarn.jar

./bin/spark-submit \
--class com.imooc.spark.Test.TestOfSparkContext2OnYarn \
--master yarn \
--deploy-mode client \
/home/hadoop/data/jar-test/sql-3.0-onYarn.jar

注意: spark-submit 提交jar到yarn,数据输入输出路径都必须是HDFS路径,否则报错 :Input path does not exist

查看
日志直接在控制台输出

结果
[hadoop@hadoop001 ~]$ hadoop fs -ls /user/hadoop/emp-spark-test/emp-jsonOnYarnTable
18/11/20 17:42:27 WARN util.NativeCodeLoader: Unable to load native-hadoop library for your platform... using builtin-java classes where applicable
Found 5 items
-rw-r--r--   1 hadoop supergroup          0 2018-11-20 17:33 /user/hadoop/emp-spark-test/emp-jsonOnYarnTable/_SUCCESS
drwxr-xr-x   - hadoop supergroup          0 2018-11-20 17:33 /user/hadoop/emp-spark-test/emp-jsonOnYarnTable/deptno=10
drwxr-xr-x   - hadoop supergroup          0 2018-11-20 17:33 /user/hadoop/emp-spark-test/emp-jsonOnYarnTable/deptno=20
drwxr-xr-x   - hadoop supergroup          0 2018-11-20 17:33 /user/hadoop/emp-spark-test/emp-jsonOnYarnTable/deptno=30
drwxr-xr-x   - hadoop supergroup          0 2018-11-20 17:33 /user/hadoop/emp-spark-test/emp-jsonOnYarnTable/deptno=44

注意
本地提交:输入输出路文件在本地写法:    file:///
输入输出 HDFS写法:                              hdfs://ip:port/
其中提交模式是yarn模式,输入输出只能是在hdfs上
yarn cluster 模式
提交命令
sql-1.0-yarnCluster.jar
./bin/spark-submit \
--class com.imooc.spark.Test.TestOfSparkContext2OnYarn \
--master yarn \
----deploy-mode cluster  \
/home/hadoop/data/jar-test/sql-3.0-onYarn.jar

查看输出日志，yarn的日志查看

Spark on YARN 之 client 与cluster 模式区别
 
  

yarn-client	Driver运行在Client端;Client请求container完成作业调度执行,client不能退出;日志在控制台输出,方便查看
yarn-cluster	Driver运行在applicationMaster;Client一旦提交作业就可以关掉,作业已经运行在yarn上;日志在客户端看不到,因为作业运行在yarn上,通过 yarn logs -applicationIdapplication_id 查看

standalone 模式
代码和local模式一样

提交
./bin/spark-submit 
–class com.imooc.spark.Test.TestOfSparkContext2 
–master spark://192.168.52.130:7077 
–executor-memory 4G 
–total-executor-cores 6 
/home/hadoop/data/jar-test/sql-1.0-yarnCluster.jar

三种模式比较
模式                   简介
local               开发使用
standalone    Spark自带,如果一个集群是Standalone的话,需要在每台机器上部署Spark
YARN            生产使用模式,统一使用YARN对整个集群作业(MR/Spark)的资源调度
Mesos    
提交参数详解
参数    作用
–class    jar包的包路径,如:com.imooc.spark.Test.TestOfSparkContext2OnYarn
–master    1.local[k]:本地模式,设置k个core ;2.yarn模式(1)-master yarn \ ----deploy-mode yarn-client–master;(2) yarn \ ----deploy-mode yarn-cluster 注意:不写----deploy-mode:默认是yarn-client模式;
执行jar包    3./home/hadoop/data/jar-test/sql-3.0-onYarn.jar 执行的jar包
–executor-memory    分配内存
–master local[8]    8个core
1000    1000个task

原文：https://blog.csdn.net/huonan_123/article/details/84282843