我正在尝试查找Spark数据帧中多列的最大值。每个列的值都为double类型。
数据框类似于:
>>> all(a > b for (a, b) in zip(A, B))
False
期望是:
+-----+---+----+---+---+
|Name | A | B | C | D |
+-----+---+----+---+---+
|Alex |5.1|-6.2| 7| 8|
|John | 7| 8.3| 1| 2|
|Alice| 5| 46| 3| 2|
|Mark |-20| -11|-22| -5|
+-----+---+----+---+---+
答案 0 :(得分:4)
您可以将greatest
应用于数字列列表,如下所示:
import org.apache.spark.sql.types._
import org.apache.spark.sql.functions._
import spark.implicits._
val df = Seq(
("Alex", 5.1, -6.2, 7.0, 8.0),
("John", 7.0, 8.3, 1.0, 2.0),
("Alice", 5.0, 46.0, 3.0, 2.0),
("Mark", -20.0, -11.0, -22.0, -5.0),
).toDF("Name", "A", "B", "C", "D")
val numCols = df.columns.tail // Apply suitable filtering as needed (*)
df.withColumn("MaxValue", greatest(numCols.head, numCols.tail: _*)).
show
// +-----+-----+-----+-----+----+--------+
// | Name| A| B| C| D|MaxValue|
// +-----+-----+-----+-----+----+--------+
// | Alex| 5.1| -6.2| 7.0| 8.0| 8.0|
// | John| 7.0| 8.3| 1.0| 2.0| 8.3|
// |Alice| 5.0| 46.0| 3.0| 2.0| 46.0|
// | Mark|-20.0|-11.0|-22.0|-5.0| -5.0|
// +-----+-----+-----+-----+----+--------+
(*)例如,要过滤所有顶级DoubleType
列:
import org.apache.spark.sql.types._
val numCols = df.schema.fields.collect{
case StructField(name, DoubleType, _, _) => name
}
如果您使用的是Spark 2.4+
,则可以选择使用array_max
,尽管在这种情况下,它会涉及附加的转换步骤:
df.withColumn("MaxValue", array_max(array(numCols.map(col): _*)))
答案 1 :(得分:-1)
首先,我复制了您的df:
scala> df.show
+-----+---+----+---+---+
| Name| A| B| C| D|
+-----+---+----+---+---+
| Alex|5.1|-6.2| 7| 8|
| John| 7| 8.3| 1| 2|
|Alice| 5| 46| 3| 2|
| Mark|-20| -11|-22| -5|
+-----+---+----+---+---+
然后我将其转换为RDD并在行级别进行转换:
import scala.math.max
case class MyData(Name: String, A: Double, B: Double, C: Double, D: Double, MaxValue: Double)
val maxDF = df.rdd.map(row => {
val a = row(1).toString.toDouble
val b = row(2).toString.toDouble
val c = row(3).toString.toDouble
val d = row(4).toString.toDouble
new MyData(row(0).toString, a, b, c, d, max(max(a, b), max(c, d)))
}).toDF
这是最终输出:
maxDF.show
+-----+-----+-----+-----+----+--------+
| Name| A| B| C| D|MaxValue|
+-----+-----+-----+-----+----+--------+
| Alex| 5.1| -6.2| 7.0| 8.0| 8.0|
| John| 7.0| 8.3| 1.0| 2.0| 8.3|
|Alice| 5.0| 46.0| 3.0| 2.0| 46.0|
| Mark|-20.0|-11.0|-22.0|-5.0| -5.0|
+-----+-----+-----+-----+----+--------+
答案 2 :(得分:-1)
您可以定义一个接收数组的UDF并返回其最大值
val getMaxColumns = udf((xs: Seq[Double]) => {
xs.max
})
然后创建要获取最大值(无论多少列)的列的数组
val columns = array($"A",$"B",$"C",$"D")
在您的示例中,由于您要应用所有尾列的最大值,因此可以
val columns = df.columns.tail.map(x => $"$x")
然后将withColumn与先前的udf一起应用
df.withColumn("maxValue", getMaxColumns(columns))
记住进口:
import org.apache.spark.sql.functions.{udf, array}
快速示例:
输入
df.show
+-----+-----+-----+-----+----+
| Name| A| B| C| D|
+-----+-----+-----+-----+----+
| Alex| 5.1| -6.2| 7.0| 8.0|
| John| 7.0| 8.3| 1.0| 2.0|
|Alice| 5.0| 46.0| 3.0| 2.0|
| Mark|-20.0|-11.0|-22.0|-5.0|
+-----+-----+-----+-----+----+
输出
df.withColumn("maxValue", getMaxColumns(columns)).show
+-----+-----+-----+-----+----+--------+
| Name| A| B| C| D|maxValue|
+-----+-----+-----+-----+----+--------+
| Alex| 5.1| -6.2| 7.0| 8.0| 8.0|
| John| 7.0| 8.3| 1.0| 2.0| 8.3|
|Alice| 5.0| 46.0| 3.0| 2.0| 46.0|
| Mark|-20.0|-11.0|-22.0|-5.0| -5.0|
+-----+-----+-----+-----+----+--------+