Apache spark 聚合后如何包含非聚合列？

我使用的是spark-sql-2.4.1v。 这里我有如下场景

val df = Seq(
  (2010,"2018-11-24",71285,"USA","0.9192019",  "0.1992019",  "0.9955999"),
  (2010,"2017-08-24",71286,"USA","0.9292018",  "0.2992019",  "0.99662018"),
  (2010,"2019-02-24",71287,"USA","0.9392017",  "0.3992019",  "0.99772000")).toDF("seq_id","load_date","company_id","country_code","item1_value","item2_value","item3_value")
.withColumn("item1_value", $"item1_value".cast(DoubleType))
.withColumn("item2_value", $"item2_value".cast(DoubleType))
.withColumn("item3_value", $"item3_value".cast(DoubleType))
.withColumn("fiscal_year", year(col("load_date")).cast(IntegerType))
.withColumn("fiscal_quarter", quarter(col("load_date")).cast(IntegerType))

df.show()


val aggregateColumns = Seq("item1_value","item2_value","item3_value")
var aggDFs = aggregateColumns.map( c => {
    df.groupBy("country_code").agg(lit(c).as("col_name"),sum(c).as("sum_of_column"))
})


var combinedDF = aggDFs.reduce(_ union _)
combinedDF.show

我得到的输出数据如下

我需要在输出中获得其他列，即“序列id”、“加载日期”和“公司id”

---

数据帧聚合操作后如何获取它们？

您可以使用窗口函数显示未聚合的列，也可以说显示每行的总和

如果有帮助，请参阅下面的代码段：

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

val df = Seq(
  (2010,"2018-11-24",71285,"USA","0.9192019",  "0.1992019",  "0.9955999"),
  (2010,"2017-08-24",71286,"USA","0.9292018",  "0.2992019",  "0.99662018"),
  (2010,"2019-02-24",71287,"USA","0.9392017",  "0.3992019",  "0.99772000")).
  toDF("seq_id","load_date","company_id","country_code","item1_value","item2_value","item3_value").
  withColumn("item1_value", $"item1_value".cast(DoubleType)).
  withColumn("item2_value", $"item2_value".cast(DoubleType)).
  withColumn("item3_value", $"item3_value".cast(DoubleType)).
  withColumn("fiscal_year", year(col("load_date")).cast(IntegerType)).
  withColumn("fiscal_quarter", quarter(col("load_date")).cast(IntegerType))


val byCountry = Window.partitionBy(col("country_code"))

val aggregateColumns = Seq("item1_value","item2_value","item3_value")
var aggDFs = aggregateColumns.map( c => {
    df.withColumn("col_name",lit(c)).withColumn("sum_country", sum(c) over byCountry)
})

var combinedDF = aggDFs.reduce(_ union _)

combinedDF.
select("seq_id","load_date","company_id","country_code","col_name","sum_country").
distinct.show(100,false)

输出如下所示：

+------+----------+----------+------------+-----------+------------------+
|seq_id|load_date |company_id|country_code|col_name   |sum_country       |
+------+----------+----------+------------+-----------+------------------+
|2010  |2019-02-24|71287     |USA         |item1_value|2.7876054         |
|2010  |2018-11-24|71285     |USA         |item1_value|2.7876054         |
|2010  |2017-08-24|71286     |USA         |item1_value|2.7876054         |
|2010  |2018-11-24|71285     |USA         |item2_value|0.8976057000000001|
|2010  |2019-02-24|71287     |USA         |item2_value|0.8976057000000001|
|2010  |2017-08-24|71286     |USA         |item2_value|0.8976057000000001|
|2010  |2019-02-24|71287     |USA         |item3_value|2.9899400800000002|
|2010  |2018-11-24|71285     |USA         |item3_value|2.9899400800000002|
|2010  |2017-08-24|71286     |USA         |item3_value|2.9899400800000002|
+------+----------+----------+------------+-----------+------------------+

---

您可以使用窗口函数显示未聚合的列，也可以说显示每行的总和

如果有帮助，请参阅下面的代码段：

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

val df = Seq(
  (2010,"2018-11-24",71285,"USA","0.9192019",  "0.1992019",  "0.9955999"),
  (2010,"2017-08-24",71286,"USA","0.9292018",  "0.2992019",  "0.99662018"),
  (2010,"2019-02-24",71287,"USA","0.9392017",  "0.3992019",  "0.99772000")).
  toDF("seq_id","load_date","company_id","country_code","item1_value","item2_value","item3_value").
  withColumn("item1_value", $"item1_value".cast(DoubleType)).
  withColumn("item2_value", $"item2_value".cast(DoubleType)).
  withColumn("item3_value", $"item3_value".cast(DoubleType)).
  withColumn("fiscal_year", year(col("load_date")).cast(IntegerType)).
  withColumn("fiscal_quarter", quarter(col("load_date")).cast(IntegerType))


val byCountry = Window.partitionBy(col("country_code"))

val aggregateColumns = Seq("item1_value","item2_value","item3_value")
var aggDFs = aggregateColumns.map( c => {
    df.withColumn("col_name",lit(c)).withColumn("sum_country", sum(c) over byCountry)
})

var combinedDF = aggDFs.reduce(_ union _)

combinedDF.
select("seq_id","load_date","company_id","country_code","col_name","sum_country").
distinct.show(100,false)

输出如下所示：

+------+----------+----------+------------+-----------+------------------+
|seq_id|load_date |company_id|country_code|col_name   |sum_country       |
+------+----------+----------+------------+-----------+------------------+
|2010  |2019-02-24|71287     |USA         |item1_value|2.7876054         |
|2010  |2018-11-24|71285     |USA         |item1_value|2.7876054         |
|2010  |2017-08-24|71286     |USA         |item1_value|2.7876054         |
|2010  |2018-11-24|71285     |USA         |item2_value|0.8976057000000001|
|2010  |2019-02-24|71287     |USA         |item2_value|0.8976057000000001|
|2010  |2017-08-24|71286     |USA         |item2_value|0.8976057000000001|
|2010  |2019-02-24|71287     |USA         |item3_value|2.9899400800000002|
|2010  |2018-11-24|71285     |USA         |item3_value|2.9899400800000002|
|2010  |2017-08-24|71286     |USA         |item3_value|2.9899400800000002|
+------+----------+----------+------------+-----------+------------------+

---

用下面的代码片段替换您的代码

scala> val W = Window.partitionBy("country_code")
scala> val aggDFs = aggregateColumns.map( c => {
     | df.withColumn("col_name", lit(c)).withColumn("sum_of_column" ,sum(c).over(W)).select("seq_id","load_date", "company_id","col_name","sum_of_column")
     | })

scala> val combinedDF = aggDFs.reduce(_ union _)
scala> combinedDF.show()
+------+----------+----------+-----------+------------------+                   
|seq_id| load_date|company_id|   col_name|     sum_of_column|
+------+----------+----------+-----------+------------------+
|  2010|2018-11-24|     71285|item1_value|         2.7876054|
|  2010|2017-08-24|     71286|item1_value|         2.7876054|
|  2010|2019-02-24|     71287|item1_value|         2.7876054|
|  2010|2018-11-24|     71285|item2_value|         0.8976057|
|  2010|2017-08-24|     71286|item2_value|         0.8976057|
|  2010|2019-02-24|     71287|item2_value|         0.8976057|
|  2010|2018-11-24|     71285|item3_value|2.9899400800000002|
|  2010|2017-08-24|     71286|item3_value|2.9899400800000002|
|  2010|2019-02-24|     71287|item3_value|2.9899400800000002|
+------+----------+----------+-----------+------------------+

---

用下面的代码片段替换您的代码

scala> val W = Window.partitionBy("country_code")
scala> val aggDFs = aggregateColumns.map( c => {
     | df.withColumn("col_name", lit(c)).withColumn("sum_of_column" ,sum(c).over(W)).select("seq_id","load_date", "company_id","col_name","sum_of_column")
     | })

scala> val combinedDF = aggDFs.reduce(_ union _)
scala> combinedDF.show()
+------+----------+----------+-----------+------------------+                   
|seq_id| load_date|company_id|   col_name|     sum_of_column|
+------+----------+----------+-----------+------------------+
|  2010|2018-11-24|     71285|item1_value|         2.7876054|
|  2010|2017-08-24|     71286|item1_value|         2.7876054|
|  2010|2019-02-24|     71287|item1_value|         2.7876054|
|  2010|2018-11-24|     71285|item2_value|         0.8976057|
|  2010|2017-08-24|     71286|item2_value|         0.8976057|
|  2010|2019-02-24|     71287|item2_value|         0.8976057|
|  2010|2018-11-24|     71285|item3_value|2.9899400800000002|
|  2010|2017-08-24|     71286|item3_value|2.9899400800000002|
|  2010|2019-02-24|     71287|item3_value|2.9899400800000002|
+------+----------+----------+-----------+------------------+

---

在获得“seq_id”和其他字段后，您的输出会是什么样子？我们正在考虑从基本数据帧进行联接吗？您可以添加所需的输出吗？您可以通过创建所需的列

item1\u值

，避免这种昂贵的

reduce（\uuunion\uuu）

（这是迭代且缓慢的），

item2\u值

等，而不是6列结果数据框，创建一个多列数据集，并在其各自的列中进行聚合。只是说……在获得“seq_id”和其他字段后，您的输出会是什么样子？我们是否考虑从基本数据帧进行联接？您可以添加所需的输出吗？您可以通过创建所需的列

item1\u value

，避免这种昂贵的

reduce（uuu-union）

（这是迭代且缓慢的），

item2\u值

等，而不是6列结果数据框，创建一个多列数据集，并在其各自的列中进行聚合。只是说…@b工程师，使用上面的代码，尝试使用val而不是var。你能告诉我这个广播变量访问有什么问题吗@b工程师，使用上面的代码，尝试使用val而不是var。您能告诉我这个广播变量访问有什么问题吗？