spark相關介紹-提取hive表（一）

本文環境說明

centos伺服器
jupyter的scala核spylon-kernel
spark-2.4.0
scala-2.11.12
hadoop-2.6.0

本文主要內容

spark讀取hive表的資料，主要包括直接sql讀取hive表；通過hdfs檔案讀取hive表，以及hive分割槽表的讀取。
通過jupyter上的cell來初始化sparksession。
文末還有通過spark提取hdfs檔案的完整示例

jupyter配置檔案

我們可以在jupyter的cell框裡面，對spark的session做出對應的初始化，具體可以見下面的示例。

%%init_spark
launcher.master = "local[*]"
launcher.conf.spark.app.name = "BDP-xw"
launcher.conf.spark.driver.cores = 2
launcher.conf.spark.num_executors = 3
launcher.conf.spark.executor.cores = 4
launcher.conf.spark.driver.memory = '4g'
launcher.conf.spark.executor.memory = '4g'

// launcher.conf.spark.serializer = "org.apache.spark.serializer.KryoSerializer"
// launcher.conf.spark.kryoserializer.buffer.max = '4g'

import org.apache.spark.sql.SparkSession
var NumExecutors = spark.conf.getOption("spark.num_executors").repr
var ExecutorMemory = spark.conf.getOption("spark.executor.memory").repr
var AppName = spark.conf.getOption("spark.app.name").repr
var max_buffer = spark.conf.getOption("spark.kryoserializer.buffer.max").repr

println(f"Config as follows: \nNumExecutors: $NumExecutors, \nAppName: $AppName,\nmax_buffer:$max_buffer")

直接檢視我們初始化的sparksession對應的環境各變數

從hive中取數

直接sparksql走起

import org.apache.spark.sql.SparkSession
val sql_1 = """select * from tbs limit 4 """
var df = sql(sql_1)
df.show(5, false)

通過hdfs取數

具體示例可以參考文末的從hdfs取數完整指令碼示例

object LoadingData_from_hdfs_base extends mylog{// with Logging
    ...
    def main(args: Array[String]=Array("tb1", "3", "\001", "cols", "")): Unit = {
          if (args.length < 2) {
          println("Usage: LoadingData_from_hdfs <tb_name, parts. sep_line, cols, paths>")
           System.err.println("Usage: LoadingData_from_hdfs <tb_name, parts, sep_line, cols, paths>")
           System.exit(1)
          }
        log.warn("開始啦排程")
        val tb_name = args(0)
        val parts = args(1)
        val sep_line = args(2)
        val select_col = args(3)
        val save_paths = args(4)
        val select_cols = select_col.split("#").toSeq
        log.warn(s"Loading cols are : \n $select_cols")
        val gb_sql = s"DESCRIBE FORMATTED ${tb_name}"
        val gb_desc = sql(gb_sql)
        val hdfs_address = gb_desc.filter($"col_name".contains("Location")).take(1)(0).getString(1)
        val hdfs_address_cha = s"$hdfs_address/*/"
        val Cs = new DataProcess_base(spark)
        val tb_desc = Cs.get_table_desc(tb_name)
        val raw_data = Cs.get_hdfs_data(hdfs_address)
        val len1 = raw_data.map(item => item.split(sep_line)).first.length
        val names = tb_desc.filter(!$"col_name".contains("#")).dropDuplicates(Seq("col_name")).sort("id").select("col_name").take(len1).map(_(0)).toSeq.map(_.toString)
        val schema1 = StructType(names.map(fieldName => StructField(fieldName, StringType)))
        val rawRDD = raw_data.map(_.split(sep_line).map(_.toString)).map(p => Row(p: _*)).filter(_.length == len1)
        val df_data = spark.createDataFrame(rawRDD, schema1)//.filter("custommsgtype = '1'")
        val df_desc = select_cols.toDF.join(tb_desc, $"value"===$"col_name", "left")
        val df_gb_result = df_data.select(select_cols.map(df_data.col(_)): _*)//.limit(100)
        df_gb_result.show(5, false)
        ...
//         spark.stop()
    }
}

val cols = "area_name#city_name#province_name"
val tb_name = "tb1"
val sep_line = "\u0001"
// 執行指令碼
LoadingData_from_hdfs_base.main(Array(tb_name, "4", sep_line, cols, ""))

)

判斷路徑是否為資料夾

方式1

def pathIsExist(spark: SparkSession, path: String): Boolean = {
    val filePath = new org.apache.hadoop.fs.Path( path )
    val fileSystem = filePath.getFileSystem( spark.sparkContext.hadoopConfiguration )
    fileSystem.exists( filePath )
}
pathIsExist(spark, hdfs_address)
// 得到結果如下：
// pathIsExist: (spark: org.apache.spark.sql.SparkSession, path: String)Boolean
// res4: Boolean = true

方式2

import java.io.File
val d = new File("/usr/local/xw")
d.isDirectory
// 得到結果如下：
// d: java.io.File = /usr/local/xw
// res3: Boolean = true

分割槽表讀取源資料

對分割槽檔案需要注意下，需要保證原始的hdfs上的raw檔案裡面是否有對應的分割槽欄位值
- 如果分割槽欄位在hdfs中的原始檔案中，則可以直接通過通過hdfs取數
- 若原始檔案中，不包括分割槽欄位資訊，則需要按照以下方式取數啦
- 具體示例可以參考文末的從hdfs取數完整指令碼示例

單個檔案讀取

object LoadingData_from_hdfs_onefile_with_path extends mylog{
    def main(args: Array[String]=Array("tb_name", "hdfs:/", "3","\n", "\001", "cols", "")): Unit = {
        ...
        val hdfs_address = args(1)
        val len1 = raw_data.map(item => item.split(sep_line)).first.length
        val rawRDD = raw_data.flatMap(line => line.split(sep_text)).map(word => (word.split(sep_line):+hdfs_address)).map(p => Row(p: _*))
        println(rawRDD.take(2))
        val names = tb_desc.filter(!$"col_name".contains("#")).dropDuplicates(Seq("col_name")).sort("id").select("col_name").take(len1).map(_(0)).toSeq.map(_.toString)
        import org.apache.spark.sql.types.StructType
        val schema1 = StructType(names.map(fieldName => StructField(fieldName, StringType)))
        val new_schema1 = schema1.add(StructField("path", StringType))
        val df_data = spark.createDataFrame(rawRDD, new_schema1)
        val df_desc = select_cols.toDF.join(tb_desc, $"value"===$"col_name", "left")
        // df_desc.show(false)
        val df_gb_result = df_data.select(select_cols.map(df_data.col(_)): _*)//.limit(100)
        df_gb_result.show(5, false)
        ...
//         spark.stop()
    }
}

val file1 = "hdfs:file1.csv"
val tb_name = "tb_name"
val sep_text = "\n"
val sep_line = "\001"
val cols = "city#province#etl_date#path"
// 執行指令碼
LoadingData_from_hdfs_onefile_with_path.main(Array(tb_name, file1, "4", sep_line, sep_text, cols, ""))

多個檔案讀取嘗試1

通過wholeTextFiles讀取資料夾裡面的檔案時，保留檔名資訊；
具體示例可以參考文末的從hdfs取數完整指令碼示例

object LoadingData_from_hdfs_wholetext_with_path extends mylog{// with Logging
    ...
    def main(args: Array[String]=Array("tb1", "hdfs:/", "3","\n", "\001", "cols", "")): Unit = {
        ...
        val tb_name = args(0)
        val hdfs_address = args(1)
        val parts = args(2)
        val sep_line = args(3)
        val sep_text = args(4)
        val select_col = args(5)
        val save_paths = args(6)
        val select_cols = select_col.split("#").toSeq
        val Cs = new DataProcess_get_data(spark)
        val tb_desc = Cs.get_table_desc(tb_name)
        val rddWhole = spark.sparkContext.wholeTextFiles(s"$hdfs_address", 10)
        rddWhole.foreach(f=>{
            println(f._1+"資料量是=>"+f._2.split("\n").length)
        })
        val files = rddWhole.collect
        val len1 = files.flatMap(item => item._2.split(sep_text)).take(1).flatMap(items=>items.split(sep_line)).length
        val names = tb_desc.filter(!$"col_name".contains("#")).dropDuplicates(Seq("col_name")).sort("id").select("col_name").take(len1).map(_(0)).toSeq.map(_.toString)
        import org.apache.spark.sql.types.StructType
        // 解析wholeTextFiles讀取的結果並轉化成dataframe
        val wordCount = files.map(f=>f._2.split(sep_text).map(g=>g.split(sep_line):+f._1.split("/").takeRight(1)(0))).flatMap(h=>h).map(p => Row(p: _*))
        val schema1 = StructType(names.map(fieldName => StructField(fieldName, StringType)))
        val new_schema1 = schema1.add(StructField("path", StringType))
        val rawRDD = sc.parallelize(wordCount)
        val df_data = spark.createDataFrame(rawRDD, new_schema1)
        val df_desc = select_cols.toDF.join(tb_desc, $"value"===$"col_name", "left")
        //df_desc.show(false)
        val df_gb_result = df_data.select(select_cols.map(df_data.col(_)): _*)
        df_gb_result.show(5, false)
        println("生成的dataframe,依path列groupby的結果如下")
        df_gb_result.groupBy("path").count().show(false)
        ...
//         spark.stop()
    }
}

val file1 = "hdfs:file1_1[01].csv"
val tb_name = "tb_name"
val sep_text = "\n"
val sep_line = "\001"
val cols = "city#province#etl_date#path"
// 執行指令碼
LoadingData_from_hdfs_wholetext_with_path.main(Array(tb_name, file1, "4", sep_line, sep_text, cols, ""))

讀取多檔案且保留檔名為列名技術實現

以下實現功能
- 將Array[(String, String)]型別的按(String, String)拆成多行；
- 將(String, String)中的第2個元素，按照\n分割符分成多行，按\?分隔符分成多列；
- 將(String, String)中的第1個元素，分別加到2中的每行後面。在dataframe中呈現的就是新增一列啦
業務場景
- 如果要一次讀取多個檔案，且相對合並後的資料集中，對資料來源於哪一個檔案作出區分。

// 測試用例，主要是把wholetextfile得到的結果轉化為DataFrame
val test1 = Array(("abasdfsdf", "a?b?c?d\nc?d?d?e"), ("sdfasdf", "b?d?a?e\nc?d?e?f"))
val test2 = test1.map(line=>line._2.split("\n").map(line1=>line1.split("\\?"):+line._1)).flatMap(line2=>line2).map(p => Row(p: _*))
val cols = "cn1#cn2#cn3#cn4#path"
val names = cols.split("#")
val schema1 = StructType(names.map(fieldName => StructField(fieldName, StringType)))
val rawRDD = sc.parallelize(test2)
val df_data = spark.createDataFrame(rawRDD, schema1)
df_data.show(4, false)
test1

多個檔案讀取for迴圈

通過for迴圈遍歷讀取資料夾裡面的檔案時，保留檔名資訊；
具體示例可以參考文末的從hdfs取數完整指令碼示例

import org.apache.spark.rdd.RDD
import org.apache.spark.sql.{DataFrame, Row}
import org.apache.spark.sql.SparkSession
import org.apache.spark.{SparkConf, SparkContext}
import org.apache.spark.sql.functions.monotonically_increasing_id
import org.apache.log4j.{Level, Logger}
import org.apache.spark.sql.types.{StructType, StructField, StringType, IntegerType}
import org.apache.hadoop.fs.{FileSystem, Path}
Logger.getLogger("org").setLevel(Level.WARN)
// val log = Logger.getLogger(this.getClass)
@transient lazy val log:Logger = Logger.getLogger(this.getClass)
class DataProcess_get_data_byfor (ss: SparkSession) extends java.io.Serializable{
  import ss.implicits._
  import ss.sql
  import org.apache.spark.sql.types.DataTypes
  ...
  def union_dataframe(df_1:RDD[String], df_2:RDD[String]):RDD[String] ={
    val count1 = df_1.map(item=>item.split(sep_line)).take(1)(0).length
    val count2 = df_2.map(item=>item.split(sep_line)).take(1)(0).length
    val name2 = df_2.name.split("/").takeRight(1)(0)
    val arr2 = df_2.map(item=>item.split(sep_line):+name2).map(p => Row(p: _*))
    println(s"執行到這兒了")
    var name1 = ""
    var arr1 = ss.sparkContext.makeRDD(List().map(p => Row(p: _*)))
//     var arr1 = Array[org.apache.spark.sql.Row]
    if (count1 == count2){
        name1 = df_1.name.split("/").takeRight(1)(0)
        arr1 = df_1.map(item=>item.split(sep_line):+name1).map(p => Row(p: _*))
        // arr1.foreach(f=>print(s"arr1嘞$f" + f.length + "\n"))
        println(s"執行到這兒了沒？$count1~$count2 $name1/$name2")
        arr1
    }
    else{
        println(s"執行到這兒了不相等哈？$count1~$count2 $name1/$name2")
        arr1 = df_1.map(item=>item.split(sep_line)).map(p => Row(p: _*))
    }
    var rawRDD = arr1.union(arr2)
    // arr3.foreach(f=>print(s"$f" + f.length + "\n"))
    // var rawRDD = sc.parallelize(arr3)
    var sourceRdd = rawRDD.map(_.mkString(sep_line))
//     var count31 = arr1.take(1)(0).length
//     var count32 = arr2.take(1)(0).length
//     var count3 = sourceRdd.map(item=>item.split(sep_line)).take(1)(0).length
//     var nums = sourceRdd.count
//     print(s"arr1: $count31、arr2: $count32、arr3: $count3, 資料量為:$nums")
    sourceRdd
}
}

object LoadingData_from_hdfs_text_with_path_byfor extends mylog{// with Logging
    ...
    def main(args: Array[String]=Array("tb1", "hdfs:/", "3","\n", "\001", "cols","data1", "test", "")): Unit = {
        ...
        val hdfs_address = args(1)
        ...
        val pattern = args(6)
        val pattern_no = args(7)
        val select_cols = select_col.split("#").toSeq
        log.warn(s"Loading cols are : \n $select_cols")
        val files = FileSystem.get(spark.sparkContext.hadoopConfiguration).listStatus(new Path(s"$hdfs_address"))
        val files_name = files.toList.map(f=> f.getPath.getName)
        val file_filter = files_name.filter(_.contains(pattern)).filterNot(_.contains(pattern_no))
        val df_1 = file_filter.map(item=> sc.textFile(s"$path1$item"))
        df_1.foreach(f=>{
            println(f + "資料量是" +  f.count)
        })
        val df2 = df_1.reduce(_ union _)
        println("合併後的資料量是" + df2.count)
        val Cs = new DataProcess_get_data_byfor(spark)
        ...
        // 將for迴圈讀取的結果合併起來
        val result = df_1.reduce((a, b)=>union_dataframe(a, b))
        val result2 = result.map(item=>item.split(sep_line)).map(p => Row(p: _*))
        val df_data = spark.createDataFrame(result2, new_schema1)
        val df_desc = select_cols.toDF.join(tb_desc, $"value"===$"col_name", "left")
        println("\n")
        //df_desc.show(false)
        val df_gb_result = df_data.select(select_cols.map(df_data.col(_)): _*)
        df_gb_result.show(5, false)
        println("生成的dataframe,依path列groupby的結果如下")
        df_gb_result.groupBy("path").count().show(false)
        ...
//         spark.stop()
    }
}

val path1 = "hdfs:202001/"
val tb_name = "tb_name"
val sep_text = "\n"
val sep_line = "\001"
val cols = "city#province#etl_date#path"
val pattern = "result_copy_1"
val pattern_no = "1.csv"
// val file_filter = List("file1_10.csv", "file_12.csv", "file_11.csv")

// 執行指令碼
LoadingData_from_hdfs_text_with_path_byfor.main(Array(tb_name, path1, "4", sep_line, sep_text, cols, pattern, pattern_no, ""))

執行指令碼的完整示例

import org.apache.spark.rdd.RDD
import org.apache.spark.sql.{DataFrame, Row}
import org.apache.spark.sql.SparkSession
import org.apache.spark.sql.functions.monotonically_increasing_id
import org.apache.log4j.{Level, Logger}
import org.apache.spark.sql.types.{StructType, StructField, StringType, IntegerType}
Logger.getLogger("org").setLevel(Level.WARN)
val log = Logger.getLogger(this.getClass)
class DataProcess_base (ss: SparkSession) extends java.io.Serializable{
  import ss.implicits._
  import ss.sql
  import org.apache.spark.sql.types.DataTypes
  def get_table_desc(tb_name:String="tb"):DataFrame ={
    val gb_sql = s"desc ${tb_name}"
    val gb_desc = sql(gb_sql)
    val names = gb_desc.filter(!$"col_name".contains("#")).withColumn("id", monotonically_increasing_id())
    names
  }
  def get_hdfs_data(hdfs_address:String="hdfs:"):RDD[String]={
      val gb_data = ss.sparkContext.textFile(hdfs_address)
      gb_data.cache()
      val counts1 = gb_data.count
      println(f"the rows of origin hdfs data is $counts1%-1d")
      gb_data
  }
}

    object LoadingData_from_hdfs_base extends mylog{// with Logging
        Logger.getLogger("org").setLevel(Level.WARN)
        val conf = new SparkConf()
        conf.setMaster("yarn")
        conf.setAppName("LoadingData_From_hdfs")
        conf.set("spark.home", System.getenv("SPARK_HOME"))
        val spark = SparkSession.builder().config(conf).enableHiveSupport().getOrCreate()
        import spark.implicits._
        import spark.sql
        var UIAddress = spark.conf.getOption("spark.driver.appUIAddress").repr
        var yarnserver = spark.conf.getOption("spark.org.apache.hadoop.yarn.server.webproxy.amfilter.AmIpFilter.param.PROXY_URI_BASES").repr
        println(f"Config as follows: \nUIAddress: $UIAddress, \nyarnserver: $yarnserver")
        def main(args: Array[String]=Array("tb1", "3", "\001", "cols", "")): Unit = {
              if (args.length < 2) {
              println("Usage: LoadingData_from_hdfs <tb_name, parts. sep_line, cols, paths>")
               System.err.println("Usage: LoadingData_from_hdfs <tb_name, parts, sep_line, cols, paths>")
               System.exit(1)
              }
            log.warn("開始啦排程")
            val tb_name = args(0)
            val parts = args(1)
            val sep_line = args(2)
            val select_col = args(3)
            val save_paths = args(4)
            val select_cols = select_col.split("#").toSeq
            log.warn(s"Loading cols are : \n $select_cols")
            val gb_sql = s"DESCRIBE FORMATTED ${tb_name}"
            val gb_desc = sql(gb_sql)
            val hdfs_address = gb_desc.filter($"col_name".contains("Location")).take(1)(0).getString(1)
            println(s"tbname路徑是$hdfs_address")
            val hdfs_address_cha = s"$hdfs_address/*/"
            val Cs = new DataProcess_base(spark)
            val tb_desc = Cs.get_table_desc(tb_name)
            val raw_data = Cs.get_hdfs_data(hdfs_address)
            val len1 = raw_data.map(item => item.split(sep_line)).first.length
            val names = tb_desc.filter(!$"col_name".contains("#")).dropDuplicates(Seq("col_name")).sort("id").select("col_name").take(len1).map(_(0)).toSeq.map(_.toString)
            val schema1 = StructType(names.map(fieldName => StructField(fieldName, StringType)))
            val rawRDD = raw_data.map(_.split(sep_line).map(_.toString)).map(p => Row(p: _*)).filter(_.length == len1)
            val df_data = spark.createDataFrame(rawRDD, schema1)//.filter("custommsgtype = '1'")
            val df_desc = select_cols.toDF.join(tb_desc, $"value"===$"col_name", "left")
            val df_gb_result = df_data.select(select_cols.map(df_data.col(_)): _*)//.limit(100)
            df_gb_result.show(5, false)
            println("生成的dataframe,依path列groupby的結果如下")
            // val part = parts.toInt
            // df_gb_result.repartition(part).write.mode("overwrite").option("header","true").option("sep","#").csv(save_paths)
            // log.warn(f"the rows of origin data compare to mysql results is $ncounts1%-1d VS $ncounts3%-4d")
            //         spark.stop()
        }
    }

val cols = "area_name#city_name#province_name"
val tb_name = "tb1"
val sep_line = "\u0001"
// 執行指令碼
LoadingData_from_hdfs_base.main(Array(tb_name, "4", sep_line, cols, ""))

spark相關介紹-提取hive表（一）

spark相關介紹-提取hive表（一）

本文環境說明

本文主要內容

jupyter配置檔案

從hive中取數

直接sparksql走起

通過hdfs取數

判斷路徑是否為資料夾

分割槽表讀取源資料

單個檔案讀取

多個檔案讀取嘗試1

讀取多檔案且保留檔名為列名技術實現

多個檔案讀取for迴圈

執行指令碼的完整示例

最新文章