SparkSQL ETL

记一次SparkSql ETL 过程

需求说明

1）input：json日志
2）ETL：根据IP解析出 省份，城市
3）stat: 地区分布指标计算，
满足条件的才算，满足条件的赋值为1，不满足的赋值为0 （如下图）
将统计结果写入MySQL中。
（就比如说这个广告请求要满足 requestmode=1 和 processnode =3 这两个条件）

代码分析

val spark = SparkSession.builder().master("local[2]").appName("LogApp").getOrCreate()
    import spark.implicits._
    val inputDF = spark.read.json("inputdata/data-test.json")
    inputDF.printSchema()

 	// ETL: 一定保留原有的数据 最完整   而且要落地  (理由：要是数据出错好重新计算）
    val newDF = inputDF.withColumn("province", MyUDF.getProvince(inputDF.col("ip")))
     .withColumn("city", MyUDF.getCity($"ip"))//自定义udf 函数
     .write.format("parquet")
     .mode(SaveMode.Overwrite)
     .save("outparquet") // 最好保存parquet格式 (spark默认就是parquet + snappy)

	 // 计算 重新去读取etl之后的数据源  
    val parquetDF = spark.read.parquet("outparquet/xxx.snappy.parquet")
    parquetDF.printSchema()
    parquetDF.show(5)
    parquetDF.createOrReplaceTempView("log")
    
	//业务SQL
val areaSQL01 = "select province,city, " +
      "sum(case when requestmode=1 and processnode >=1 then 1 else 0 end) origin_request," +
      "sum(case when requestmode=1 and processnode >=2 then 1 else 0 end) valid_request," +
      "sum(case when requestmode=1 and processnode =3 then 1 else 0 end) ad_request," +
      "sum(case when adplatformproviderid>=100000 and iseffective=1 and isbilling=1 and isbid=1 and adorderid!=0 then 1 else 0 end) bid_cnt," +
      "sum(case when adplatformproviderid>=100000 and iseffective=1 and isbilling=1 and iswin=1 then 1 else 0 end) bid_success_cnt," +
      "sum(case when requestmode=2 and iseffective=1 then 1 else 0 end) ad_display_cnt," +
      "sum(case when requestmode=3 and processnode=1 then 1 else 0 end) ad_click_cnt," +
      "sum(case when requestmode=2 and iseffective=1 and isbilling=1 then 1 else 0 end) medium_display_cnt," +
      "sum(case when requestmode=3 and iseffective=1 and isbilling=1 then 1 else 0 end) medium_click_cnt," +
      "sum(case when adplatformproviderid>=100000 and iseffective=1 and isbilling=1 and iswin=1 and adorderid>20000  then 1*winprice/1000 else 0 end) ad_consumption," +
      "sum(case when adplatformproviderid>=100000 and iseffective=1 and isbilling=1 and iswin=1 and adorderid>20000  then 1*adpayment/1000 else 0 end) ad_cost " +
      "from log group by province,city"
    spark.sql(areaSQL01).createOrReplaceTempView("area_tmp")

    val areaSQL02 = "select province,city, " +
      "origin_request," +
      "valid_request," +
      "ad_request," +
      "bid_cnt," +
      "bid_success_cnt," +
      "bid_success_cnt/bid_cnt bid_success_rate," +
      "ad_display_cnt," +
      "ad_click_cnt," +
      "ad_click_cnt/ad_display_cnt ad_click_rate," +
      "ad_consumption," +
      "ad_cost from area_tmp " +
      "where bid_cnt!=0 and ad_display_cnt!=0"

	// 写入MySQL （上一篇博客有介绍）
    val config = ConfigFactory.load()
    val url = config.getString("db.default.url")
    val user = config.getString("db.default.user")
    val password = config.getString("db.default.password")   

    spark.sql(areaSQL02)
      .write.format("jdbc")
      .option("url", url)
      .option("dbtable", "sparksql_test")
      .option("user", user)
      .option("password", password)
      .mode(SaveMode.Overwrite)
      .save()

       spark.stop()

自定义udf 函数代码

object MyUDF {

  import org.apache.spark.sql.functions._

  def getProvince = udf((ip:String)=>{
    val cityInfo = IPUtil.getCityInfo(ip)
    val splits = cityInfo.split("\\|")
    var city = "未知"
    if (splits.length == 5){
      city = splits(2)
    }
    city

  })


  def getCity = udf((ip:String)=>{
    val cityInfo = IPUtil.getCityInfo(ip)
    val splits = cityInfo.split("\\|")
    var city = "未知"
    if (splits.length == 5){
      city = splits(3)
    }
    city
  })

}

调优

① ETL 落地过程中应该调用coalesce() 防止产生多个小文件

 val newDF = inputDF.withColumn("province", MyUDF.getProvince(inputDF.col("ip")))
      .withColumn("city", MyUDF.getCity($"ip"))
      .coalesce(1)
     .write.format("parquet").mode(SaveMode.Overwrite).save("outparquet")

② spark.conf.set(“spark.sql.shuffle.partitions”,“400”) 修改SparkSql shuffle task数量，默认是200

总结

ETL过程：
input:json
清洗 ==> ODS 大宽表 HDFS/Hive/SparkSQL
output: 列式存储 ORC/Parquet （列式存储） （为啥要用这两种？ 因为ETL清洗出来的是全字段，我们不可能使用到全部字段，所以采用列式存储，用到几列就获取几列，这样就能减少I/O，性能大大提升）

Stat
==> 一个非常简单的SQL搞定
==> 复杂：多个SQL 或者 一个复杂SQL搞定

列式：ORC/Parquet
特点：把每一列的数据存放在一起
优点：减少IO 需要哪几列就直接获取哪几列
缺点：如果你还是要获取每一行中的所有列，那么性能比行式的差

行式：MySQL
一条记录有多个列 一行数据是存储在一起的
优点：
你每次查询都使用到所有的列
缺点：
大宽表有N多列，但是我们仅仅使用其中几列