版本說明：Spark-2.3.0

使用Spark SQL在對數據進行處理的過程中，可能會遇到對一列數據拆分為多列，或者把多列數據合並為一列。這裏記錄一下目前想到的對DataFrame列數據進行合並和拆分的幾種方法。

1 DataFrame列數據的合並
例如：我們有如下數據，想要將三列數據合並為一列，並以“,”分割

+----+---+-----------+
|name|age|      phone|
+----+---+-----------+
|Ming| 20|15552211521|
|hong| 19|13287994007|
| zhi| 21|15552211523|
+----+---+-----------+
 

1.1 使用map方法重寫

使用map方法重寫就是將DataFrame使用map取值之後，然後使用toSeq方法轉成Seq格式，最後使用Seq的foldLeft方法拼接數據，並返回，如下所示：

//方法1：利用map重寫
    val separator = ","
    df.map(_.toSeq.foldLeft("")(_ + separator + _).substring(1)).show()

    /**
      * +-------------------+
      * |              value|
      * +-------------------+
      * |Ming,20,15552211521|
      * |hong,19,13287994007|
      * | zhi,21,15552211523|
      * +-------------------+
      
 
*/

1.2 使用內置函數concat_ws

合並多列數據也可以使用SparkSQL的內置函數concat_ws()

//方法2： 使用內置函數 concat_ws
    import org.apache.spark.sql.functions._
    df.select(concat_ws(separator, $"name", $"age", $"phone").cast(StringType).as("value")).show()

    /**
      * +-------------------+
      * |              value|
      * +-------------------+
      * |Ming,20,15552211521|
      * |hong,19,13287994007|
      * | zhi,21,15552211523|
      * +-------------------+
      
 
*/

1.3 使用自定義UDF函數

自己編寫UDF函數，實現多列合並

 //方法3：使用自定義UDF函數

    // 編寫udf函數
    def mergeCols(row: Row): String = {
      row.toSeq.foldLeft("")(_ + separator + _).substring(1)
    }

    val mergeColsUDF = udf(mergeCols _)
    df.select(mergeColsUDF(struct($"name", $"age", $"phone")).as("value")).show()

完整代碼：

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

/**
  * Created by shirukai on 2018/9/12
  * DataFrame 合並列
  */
object MergeColsTest {
  def main(args: Array[String]): Unit = {
    val spark = SparkSession
      .builder()
      .appName(this.getClass.getSimpleName)
      .master("local")
      .getOrCreate()

    //從內存創建一組DataFrame數據
    import spark.implicits._
    val df = Seq(("Ming", 20, 15552211521L), ("hong", 19, 13287994007L), ("zhi", 21, 15552211523L))
      .toDF("name", "age", "phone")
    df.show()
    /**
      * +----+---+-----------+
      * |name|age|      phone|
      * +----+---+-----------+
      * |Ming| 20|15552211521|
      * |hong| 19|13287994007|
      * | zhi| 21|15552211523|
      * +----+---+-----------+
      */
    //方法1：利用map重寫
    val separator = ","
    df.map(_.toSeq.foldLeft("")(_ + separator + _).substring(1)).show()

    /**
      * +-------------------+
      * |              value|
      * +-------------------+
      * |Ming,20,15552211521|
      * |hong,19,13287994007|
      * | zhi,21,15552211523|
      * +-------------------+
      */
    //方法2： 使用內置函數 concat_ws
    import org.apache.spark.sql.functions._
    df.select(concat_ws(separator, $"name", $"age", $"phone").cast(StringType).as("value")).show()

    /**
      * +-------------------+
      * |              value|
      * +-------------------+
      * |Ming,20,15552211521|
      * |hong,19,13287994007|
      * | zhi,21,15552211523|
      * +-------------------+
      */
    //方法3：使用自定義UDF函數

    // 編寫udf函數
    def mergeCols(row: Row): String = {
      row.toSeq.foldLeft("")(_ + separator + _).substring(1)
    }

    val mergeColsUDF = udf(mergeCols _)
    df.select(mergeColsUDF(struct($"name", $"age", $"phone")).as("value")).show()

    /**
      * /**
      * * +-------------------+
      * * |              value|
      * * +-------------------+
      * * |Ming,20,15552211521|
      * * |hong,19,13287994007|
      * * | zhi,21,15552211523|
      * * +-------------------+
      **/
      */
  }
}

2 DataFrame列數據的拆分

上面我們將DataFrame的多列數據合並為一列如下所示，有時候我們也需要將單列數據，以某種拆分規則，拆分為多列。下面提供幾種將一列拆分為多列的方法。

+-------------------+
|              value|
+-------------------+
|Ming,20,15552211521|
|hong,19,13287994007|
| zhi,21,15552211523|
+-------------------+

2.1 使用內置函數split，然後遍歷添加列

該方法，先利用內置函數split將單列的數據拆分，然後遍歷使用getItem(角標)方法獲取拆分後的數據，依次使用withColumn方法添加新列，代碼如下所示：

  //方法1： 使用內置函數split，然後遍歷添加列
    val separator = ","
    lazy val first = df.first()

    val numAttrs = first.toString().split(separator).length
    val attrs = Array.tabulate(numAttrs)(n => "col_" + n)
    //按指定分隔符拆分value列，生成splitCols列
    var newDF = df.withColumn("splitCols", split($"value", separator))
    attrs.zipWithIndex.foreach(x => {
      newDF = newDF.withColumn(x._1, $"splitCols".getItem(x._2))
    })
    newDF.show()
  /**
      * +-------------------+--------------------+-----+-----+-----------+
      * |              value|           splitCols|col_0|col_1|      col_2|
      * +-------------------+--------------------+-----+-----+-----------+
      * |Ming,20,15552211521|[Ming, 20, 155522...| Ming|   20|15552211521|
      * |hong,19,13287994007|[hong, 19, 132879...| hong|   19|13287994007|
      * | zhi,21,15552211523|[zhi, 21, 1555221...|  zhi|   21|15552211523|
      * +-------------------+--------------------+-----+-----+-----------+

2.2 使用UDF函數創建多列數據，然後合並
該方法是使用udf函數，生成多個列，然後合並到原來的數據。該方法參考了VectorDisassembler（與spark ml官網提供的VectorAssembler相反），這是一個第三方的spark ml向量拆分算法，該方法github地址：https://github.com/jamesbconner/VectorDisassembler。代碼如下所示：

//方法2：使用udf函數創建多列，然後合並
    val attributes: Array[Attribute] = {
      val numAttrs = first.toString().split(separator).length
      //生成attributes
      Array.tabulate(numAttrs)(i => NumericAttribute.defaultAttr.withName("value" + "_" + i))
    }
    //創建多列數據
    val fieldCols = attributes.zipWithIndex.map(x => {
      val assembleFunc = udf {
        str: String =>
          str.split(separator)(x._2)
      }
      assembleFunc(df("value").cast(StringType)).as(x._1.name.get, x._1.toMetadata())
    })
    //合並數據
    df.select(col("*") +: fieldCols: _*).show()

    /**
      * +-------------------+-------+-------+-----------+
      * |              value|value_0|value_1|    value_2|
      * +-------------------+-------+-------+-----------+
      * |Ming,20,15552211521|   Ming|     20|15552211521|
      * |hong,19,13287994007|   hong|     19|13287994007|
      * | zhi,21,15552211523|    zhi|     21|15552211523|
      * +-------------------+-------+-------+-----------+
      */

完整代碼：

import org.apache.spark.ml.attribute.{Attribute, NumericAttribute}
import org.apache.spark.sql.SparkSession
import org.apache.spark.sql.types.StringType

/**
  * Created by shirukai on 2018/9/12
  * 拆分列
  */
object SplitColTest {
  def main(args: Array[String]): Unit = {
    val spark = SparkSession
      .builder()
      .appName(this.getClass.getSimpleName)
      .master("local")
      .getOrCreate()

    //從內存中創建DataFrame
    import spark.implicits._
    val df = Seq("Ming,20,15552211521", "hong,19,13287994007", "zhi,21,15552211523")
      .toDF("value")
    df.show()

    /**
      * +-------------------+
      * |              value|
      * +-------------------+
      * |Ming,20,15552211521|
      * |hong,19,13287994007|
      * | zhi,21,15552211523|
      * +-------------------+
      */

    import org.apache.spark.sql.functions._
    //方法1： 使用內置函數split，然後遍歷添加列
    val separator = ","
    lazy val first = df.first()

    val numAttrs = first.toString().split(separator).length
    val attrs = Array.tabulate(numAttrs)(n => "col_" + n)
    //按指定分隔符拆分value列，生成splitCols列
    var newDF = df.withColumn("splitCols", split($"value", separator))
    attrs.zipWithIndex.foreach(x => {
      newDF = newDF.withColumn(x._1, $"splitCols".getItem(x._2))
    })
    newDF.show()

    /**
      * +-------------------+--------------------+-----+-----+-----------+
      * |              value|           splitCols|col_0|col_1|      col_2|
      * +-------------------+--------------------+-----+-----+-----------+
      * |Ming,20,15552211521|[Ming, 20, 155522...| Ming|   20|15552211521|
      * |hong,19,13287994007|[hong, 19, 132879...| hong|   19|13287994007|
      * | zhi,21,15552211523|[zhi, 21, 1555221...|  zhi|   21|15552211523|
      * +-------------------+--------------------+-----+-----+-----------+
      */

    //方法2：使用udf函數創建多列，然後合並
    val attributes: Array[Attribute] = {
      val numAttrs = first.toString().split(separator).length
      //生成attributes
      Array.tabulate(numAttrs)(i => NumericAttribute.defaultAttr.withName("value" + "_" + i))
    }
    //創建多列數據
    val fieldCols = attributes.zipWithIndex.map(x => {
      val assembleFunc = udf {
        str: String =>
          str.split(separator)(x._2)
      }
      assembleFunc(df("value").cast(StringType)).as(x._1.name.get, x._1.toMetadata())
    })
    //合並數據
    df.select(col("*") +: fieldCols: _*).show()

    /**
      * +-------------------+-------+-------+-----------+
      * |              value|value_0|value_1|    value_2|
      * +-------------------+-------+-------+-----------+
      * |Ming,20,15552211521|   Ming|     20|15552211521|
      * |hong,19,13287994007|   hong|     19|13287994007|
      * | zhi,21,15552211523|    zhi|     21|15552211523|
      * +-------------------+-------+-------+-----------+
      */
  }
}

Spark DataFrame列的合並與拆分