spark操作总结
一、sparkContext与sparkSession区别
任何Spark程序都是SparkContext开始的,SparkContext的初始化需要一个SparkConf对象,SparkConf包含了Spark集群配置的各种参数,sparkContext只能在driver机器上面启动;
SparkSession: SparkSession实质上是SQLContext和HiveContext的组合,SparkSession内部封装了sparkContext,所以计算实际上是由sparkContext完成
val conf: SparkConf = new SparkConf().setAppName("test") val spark: SparkSession = SparkSession.builder().config(conf).enableHiveSupport().getOrCreate()
二、repartition与coalesce区别
repartition一般是用来增加分区数(当然也可以减少),coalesce只能用来减少分区数。所以如果不介意保存的文件块大小不一样,可以使用coalesce来减少分区数,保存的时候一个分区就会生成一个文件块
三、Scala常用方法
1. StringBuilder
主要用于字符串的拼接,可作用于生成倒排序列,如: val userItemScore = sc.parallelize(List((1001, 1, 0.8), (1001, 2, 0.7), (1001, 3, 0.5), (1001, 4, 0.9))) userItemScore.map(x => (x._1, (x._2.toString, x._3.toString))).groupByKey() .map{x => val userid = x._1 val item_score_list = x._2 val tmp_arr = item_score_list.toArray.sortWith(_._2 > _._2) val watch_len = tmp_arr.length val strbuf = new StringBuilder() for (i <- 0 until watch_len - 1) { strbuf ++= tmp_arr(i)._1 strbuf.append(":") strbuf ++= tmp_arr(i)._2 strbuf.append(" ") } strbuf ++= tmp_arr(watch_len - 1)._1 strbuf.append(":") strbuf ++= tmp_arr(watch_len - 1)._2 userid + "\t" + strbuf }.collect()
2. scala.collection.mutable.ArrayBuffer
相当于是一个大小可变数组,把需要的值添加进来,例如: val tmpArray = new ArrayBuffer[String]() val tmpArray = new ArrayBuffer[Int]() val tmpArray = new ArrayBuffer[(String, Int)]() scala> tmpArray.append(("wangzai", 1)) scala> tmpArray res11: scala.collection.mutable.ArrayBuffer[(String, Int)] = ArrayBuffer((wangzai,1), (test,2)) tmpArray.indexOf(("test",2))为获取当前值的索引,返回类型为整型 tmpArray.slice(tmpArray.indexOf(("test", 2)), tmpArray.length)为切片,返回类型为ArrayBuffer
四、通过spark-shell来操作数据库中的表
1 启动(通过--jars指定包,后面reids包不需要,只是演示添加多个包的用法)
/xxx/spark/bin/spark-shell \ --master spark://xxx:7077 \ --executor-cores 1 \ --total-executor-cores 5 \ --driver-memory 2g \ --jars /xxx/jars/mysql-connector-java-5.1.38.jar,/xxx/jars/jedis-2.9.0.jar
2 在命令行中输入::paste, 然后粘贴以下代码,最后ctrl+D退出之后,即可执行
import java.util.Properties
import org.apache.spark.sql.{DataFrame, SparkSession}
import org.apache.spark.SparkConf
val conf: SparkConf = new SparkConf()
val spark: SparkSession = SparkSession.builder().config(conf).getOrCreate()
val mysqlUrl: String = "jdbc:mysql://ip:port/database?useUnicode=true&characterEncoding=UTF-8&useSSL=false"
val productTable: String = "product_info"
val orderTable: String = "order_info"
val properties: Properties = new Properties()
properties.put("user", user)
properties.put("password", password)
// 获取同事购配置表数据
val productDF: DataFrame = spark.read.jdbc(mysqlUrl, productTable, properties).select("id", "name")
val orderDF: DataFrame = spark.read.jdbc(mysqlUrl, orderTable, properties).select("product_id", "createTime")
val totalDataDF = productDF.join(orderDF, orderDF("product_id") === productDF("id")).drop("id")
//如果product_info对应的id为product_id,即关联id字段名不相同
//val totalDataDF = productDF.join(orderDF, Seq("product_id"))
3 把该DateFrame注册为临时表才能通过spark-sql操作
totalDataDF.createOrReplaceTempView("totalDataDF")
五、spark-sql的基本操作
//默认显示20条数据 scala> df.show() //打印模式信息 scala> df.printSchema() //选择多列 scala> df.select(df("name"),df("age")+1).show() // 条件过滤 scala> df.filter(df("age") > 20 ).show() // 分组聚合 scala> df.groupBy("age").count().show() // 排序 scala> df.sort(df("age").desc).show() //多列排序 scala> df.sort(df("age").desc, df("name").asc).show() //对列进行重命名 scala> df.select(df("name").as("username"),df("age")).show()
//对多个列重命名
scala> df.withColumnRenamed("id", "userId").withColumnRenamed("name", "userName")