在Spark中获取组的最后一个值

Question

我有一个SparkR DataFrame，如下所示：

#Create R data.frame
custId <- c(rep(1001, 5), rep(1002, 3), 1003)
date <- c('2013-08-01','2014-01-01','2014-02-01','2014-03-01','2014-04-01','2014-02-01','2014-03-01','2014-04-01','2014-04-01')
desc <- c('New','New','Good','New', 'Bad','New','Good','Good','New')
newcust <- c(1,1,0,1,0,1,0,0,1)
df <- data.frame(custId, date, desc, newcust)

#Create SparkR DataFrame    
df <- createDataFrame(df)
display(df)
      custId|    date   | desc | newcust
      --------------------------------------
       1001 | 2013-08-01| New  |   1
       1001 | 2014-01-01| New  |   1
       1001 | 2014-02-01| Good |   0
       1001 | 2014-03-01| New  |   1
       1001 | 2014-04-01| Bad  |   0
       1002 | 2014-02-01| New  |   1
       1002 | 2014-03-01| Good |   0
       1002 | 2014-04-01| Good |   0 
       1003 | 2014-04-01| New  |   1

newcust表示每次新的custId出现时的新客户，或者如果相同的custId的desc恢复为“新”。我想要获得的是每组desc的最后newcust值，同时保持每个分组的第一个date。下面是我想要获取的DataFrame。我怎么能在Spark中做到这一点？ PySpark或SparkR代码都可以使用。

#What I want 
custId|    date   | newcust | finaldesc
----------------------------------------------
 1001 | 2013-08-01|   1     | New
 1001 | 2014-01-01|   1     | Good
 1001 | 2014-03-01|   1     | Bad
 1002 | 2014-02-01|   1     | Good
 1003 | 2014-04-01|   1     | New

Answer 1

我不知道sparkR所以我会在pyspark回答。您可以使用窗口函数来实现此目的。

首先，让我们定义“newcust的分组”，你希望newcust等于1的每一行都是一个新组的开始，计算一个累积和就可以了：

from pyspark.sql import Window
import pyspark.sql.functions as psf

w1 = Window.partitionBy("custId").orderBy("date")
df1 = df.withColumn("subgroup", psf.sum("newcust").over(w1))

+------+----------+----+-------+--------+
|custId|      date|desc|newcust|subgroup|
+------+----------+----+-------+--------+
|  1001|2013-08-01| New|      1|       1|
|  1001|2014-01-01| New|      1|       2|
|  1001|2014-02-01|Good|      0|       2|
|  1001|2014-03-01| New|      1|       3|
|  1001|2014-04-01| Bad|      0|       3|
|  1002|2014-02-01| New|      1|       1|
|  1002|2014-03-01|Good|      0|       1|
|  1002|2014-04-01|Good|      0|       1|
|  1003|2014-04-01| New|      1|       1|
+------+----------+----+-------+--------+

对于每个subgroup，我们希望保留第一个日期：

w2 = Window.partitionBy("custId", "subgroup")
df2 = df1.withColumn("first_date", psf.min("date").over(w2))

+------+----------+----+-------+--------+----------+
|custId|      date|desc|newcust|subgroup|first_date|
+------+----------+----+-------+--------+----------+
|  1001|2013-08-01| New|      1|       1|2013-08-01|
|  1001|2014-01-01| New|      1|       2|2014-01-01|
|  1001|2014-02-01|Good|      0|       2|2014-01-01|
|  1001|2014-03-01| New|      1|       3|2014-03-01|
|  1001|2014-04-01| Bad|      0|       3|2014-03-01|
|  1002|2014-02-01| New|      1|       1|2014-02-01|
|  1002|2014-03-01|Good|      0|       1|2014-02-01|
|  1002|2014-04-01|Good|      0|       1|2014-02-01|
|  1003|2014-04-01| New|      1|       1|2014-04-01|
+------+----------+----+-------+--------+----------+

最后，我们要保留每个subgroup的最后一行（按日期排序）：

w3 = Window.partitionBy("custId", "subgroup").orderBy(psf.desc("date"))
df3 = df2.withColumn(
    "rn", 
    psf.row_number().over(w3)
).filter("rn = 1").select(
    "custId", 
    psf.col("first_date").alias("date"), 
    "desc"
)

+------+----------+----+
|custId|      date|desc|
+------+----------+----+
|  1001|2013-08-01| New|
|  1001|2014-01-01|Good|
|  1001|2014-03-01| Bad|
|  1002|2014-02-01|Good|
|  1003|2014-04-01| New|
+------+----------+----+

Answer 2

这是@ MaFF在SparkR中的代码：

w1 <- orderBy(windowPartitionBy('custId'), df$date)
df1 <- withColumn(df, "subgroup", over(sum(df$newcust), w1))

w2 <- windowPartitionBy("custId", "subgroup")
df2 <- withColumn(df1, "first_date", over(min(df1$date), w2))

w3 <- orderBy(windowPartitionBy("custId", "subgroup"), desc(df$date))
df3 <- withColumn(df2, "rn", over(row_number(), w3))
df3 <- select(filter(df3, df3$rn == 1), "custId", "first_date", "desc")
df3 <- withColumnRenamed(df3, 'first_date', "date")

df3 <- arrange(df3, 'custId', 'date')
display(df3)
+------+----------+----+
|custId|      date|desc|
+------+----------+----+
|  1001|2013-08-01| New|
|  1001|2014-01-01|Good|
|  1001|2014-03-01| Bad|
|  1002|2014-02-01|Good|
|  1003|2014-04-01| New|
+------+----------+----+