如何在Spark SQL中的两列上聚合

Question

现在，我有一个执行以下任务的表：

1. 按部门编号和员工编号的功能分组
2. 在每个组中，我需要按（ArrivalDate，ArrivalTime）对其进行排序，然后选择第一个。 因此，如果两个日期不同，请选择较新的日期。 如果两个日期相同，则选择较新的时间。

我正在尝试这种方法：

input.select("DepartmenId","EmolyeeID", "ArrivalDate", "ArrivalTime", "Word")
  .agg(here will be the function that handles logic from 2)
  .show()

这里要聚合的语法是什么？

先感谢您。

 // +-----------+---------+-----------+-----------+--------+ // |DepartmenId|EmolyeeID|ArrivalDate|ArrivalTime| Word | // +-----------+---------+-----------+-----------+--------+ // | D1 | E1 | 20170101 | 0730 | "YES" | // +-----------+---------+-----------+-----------+--------+ // | D1 | E1 | 20170102 | 1530 | "NO" | // +-----------+---------+-----------+-----------+--------+ // | D1 | E2 | 20170101 | 0730 | "ZOO" | // +-----------+---------+-----------+-----------+--------+ // | D1 | E2 | 20170102 | 0330 | "BOO" | // +-----------+---------+-----------+-----------+--------+ // | D2 | E1 | 20170101 | 0730 | "LOL" | // +-----------+---------+-----------+-----------+--------+ // | D2 | E1 | 20170101 | 1830 | "ATT" | // +-----------+---------+-----------+-----------+--------+ // | D2 | E2 | 20170105 | 1430 | "UNI" | // +-----------+---------+-----------+-----------+--------+ // output should be // +-----------+---------+-----------+-----------+--------+ // |DepartmenId|EmolyeeID|ArrivalDate|ArrivalTime| Word | // +-----------+---------+-----------+-----------+--------+ // | D1 | E1 | 20170102 | 1530 | "NO" | // +-----------+---------+-----------+-----------+--------+ // | D1 | E2 | 20170102 | 0330 | "BOO" | // +-----------+---------+-----------+-----------+--------+ // | D2 | E1 | 20170101 | 1830 | "ATT" | // +-----------+---------+-----------+-----------+--------+ // | D2 | E2 | 20170105 | 1430 | "UNI" | // +-----------+---------+-----------+-----------+--------+ 

Answer 1

一种方法是使用“火花窗口”功能：

val df = Seq(
  ("D1", "E1", "20170101", "0730", "YES"),
  ("D1", "E1", "20170102", "1530", "NO"),
  ("D1", "E2", "20170101", "0730", "ZOO"),
  ("D1", "E2", "20170102", "0330", "BOO"),
  ("D2", "E1", "20170101", "0730", "LOL"),
  ("D2", "E1", "20170101", "1830", "ATT"),
  ("D2", "E2", "20170105", "1430", "UNI")
).toDF(
  "DepartmenId", "EmolyeeID", "ArrivalDate", "ArrivalTime", "Word"
)

import org.apache.spark.sql.expressions.Window

val df2 = df.withColumn("rowNum", row_number().over(
    Window.partitionBy("DepartmenId", "EmolyeeID").
      orderBy($"ArrivalDate".desc, $"ArrivalTime".desc)
  )).
  select("DepartmenId", "EmolyeeID", "ArrivalDate", "ArrivalTime","Word").
  where($"rowNum" === 1).
  orderBy("DepartmenId", "EmolyeeID")

df2.show
+-----------+---------+-----------+-----------+----+
|DepartmenId|EmolyeeID|ArrivalDate|ArrivalTime|Word|
+-----------+---------+-----------+-----------+----+
|         D1|       E1|   20170102|       1530|  NO|
|         D1|       E2|   20170102|       0330| BOO|
|         D2|       E1|   20170101|       1830| ATT|
|         D2|       E2|   20170105|       1430| UNI|
+-----------+---------+-----------+-----------+----+

Answer 2

您可以在包含所有非分组列的新Struct列上使用max ，首先使用ArrivalData然后使用ArrivalTime ：新列的排序将满足您的要求（最新日期为第一；相似日期中的最后一小时为第一），因此maximum将产生您想要的记录。

然后，您可以使用select操作将结构“拆分”回单独的列。

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

df.groupBy($"DepartmentID", $"EmployeeID")
  .agg(max(struct("ArrivalDate", "ArrivalTime", "Word")) as "struct")
  .select($"DepartmentID", $"EmployeeID",
    $"struct.ArrivalDate" as "ArrivalDate",
    $"struct.ArrivalTime" as "ArrivalTime",
    $"struct.Word" as "Word"
  )