Spark Scala:使用分析函数获得累计和(运行总计)

Question

我正在使用Window函数在Spark中实现累积和.但是在应用窗口分区功能时不保持记录输入的顺序

输入数据:

val base = List(List("10", "MILLER", "1300", "2017-11-03"), List("10", "Clark", "2450", "2017-12-9"), List("10", "King", "5000", "2018-01-28"),
  List("30", "James", "950", "2017-10-18"), List("30", "Martin", "1250", "2017-11-21"), List("30", "Ward", "1250", "2018-02-05"))
  .map(row => (row(0), row(1), row(2), row(3)))

val DS1 = base.toDF("dept_no", "emp_name", "sal", "date")
DS1.show()

+-------+--------+----+----------+
|dept_no|emp_name| sal|      date|
+-------+--------+----+----------+
|     10|  MILLER|1300|2017-11-03|
|     10|   Clark|2450| 2017-12-9|
|     10|    King|5000|2018-01-28|
|     30|   James| 950|2017-10-18|
|     30|  Martin|1250|2017-11-21|
|     30|    Ward|1250|2018-02-05|
+-------+--------+----+----------+

预期产出:

+-------+--------+----+----------+-----------+
|dept_no|emp_name| sal|      date|Dept_CumSal|
+-------+--------+----+----------+-----------+
|     10|  MILLER|1300|2017-11-03|     1300.0|
|     10|   Clark|2450| 2017-12-9|     3750.0|
|     10|    King|5000|2018-01-28|     8750.0|
|     30|   James| 950|2017-10-18|      950.0|
|     30|  Martin|1250|2017-11-21|     2200.0|
|     30|    Ward|1250|2018-02-05|     3450.0|
+-------+--------+----+----------+-----------+

我尝试过以下逻辑

val baseDepCumSal = DS1.withColumn("Dept_CumSal", sum("sal").over(Window.partitionBy("dept_no").
  orderBy(col("sal"), col("emp_name"), col("date").asc).
  rowsBetween(Long.MinValue, 0)
))

baseDepCumSal.orderBy("dept_no", "date").show

+-------+--------+----+----------+-----------+
|dept_no|emp_name| sal|      date|Dept_CumSal|
+-------+--------+----+----------+-----------+
|     10|  MILLER|1300|2017-11-03|     1300.0|
|     10|   Clark|2450| 2017-12-9|     3750.0|
|     10|    King|5000|2018-01-28|     8750.0|
|     30|   James| 950|2017-10-18|     3450.0|
|     30|  Martin|1250|2017-11-21|     1250.0|
|     30|    Ward|1250|2018-02-05|     2500.0|
+-------+--------+----+----------+-----------+

对于dept_no = 10,记录按预期顺序计算,而对于dept_no = 30,记录不按输入顺序计算.

Answer 1

发生这种情况是由于类型不正确。因为薪水是string

DS1.printSchema
root
 |-- dept_no: string (nullable = true)
 |-- emp_name: string (nullable = true)
 |-- sal: string (nullable = true)
 |-- date: string (nullable = true)

它按字典顺序排序：

DS1.orderBy("sal").show
+-------+--------+----+----------+
|dept_no|emp_name| sal|      date|
+-------+--------+----+----------+
|     30|  Martin|1250|2017-11-21|
|     30|    Ward|1250|2018-02-05|
|     10|  MILLER|1300|2017-11-03|
|     10|   Clark|2450| 2017-12-9|
|     10|    King|5000|2018-01-28|
|     30|   James| 950|2017-10-18|
+-------+--------+----+----------+ 

为了获得理想的结果，您必须进行强制转换（并且不需要定义框架）：

DS1.withColumn("Dept_CumSal", sum("sal").over(
  Window
     .partitionBy("dept_no")
    .orderBy(col("sal").cast("integer"), col("emp_name"), col("date").asc))).show
+-------+--------+----+----------+-----------+                                  
|dept_no|emp_name| sal|      date|Dept_CumSal|
+-------+--------+----+----------+-----------+
|     30|   James| 950|2017-10-18|      950.0|
|     30|  Martin|1250|2017-11-21|     2200.0|
|     30|    Ward|1250|2018-02-05|     3450.0|
|     10|  MILLER|1300|2017-11-03|     1300.0|
|     10|   Clark|2450| 2017-12-9|     3750.0|
|     10|    King|5000|2018-01-28|     8750.0|
+-------+--------+----+----------+-----------+

Answer 2

请注意，窗口内的顺序(col("sal"), col("emp_name"), col("date").asc)与显示的顺序不同。"dept_no", "date" 为什么窗口中需要“sal”和“emp_name”？为什么不直接按日期订购？