我有一个PySpark数据框架，它可以跟踪产品价格和状态在几个月内发生的变化。这意味着只有当与前一个月相比发生变化(状态或价格(时，才会创建新行，就像下面的伪数据一样
----------------------------------------
|product_id| status | price| month |
----------------------------------------
|1 | available | 5 | 2019-10|
----------------------------------------
|1 | available | 8 | 2020-08|
----------------------------------------
|1 | limited | 8 | 2020-10|
----------------------------------------
|2 | limited | 1 | 2020-09|
----------------------------------------
|2 | limited | 3 | 2020-10|
----------------------------------------
我想创建一个数据帧，显示过去6个月中每个月的值。这意味着，每当上述数据帧中出现间隙时，我都需要复制记录。例如，如果最后6个月是2020-07、2020-08。。。2020-12，则上述数据帧的结果应该是
----------------------------------------
|product_id| status | price| month |
----------------------------------------
|1 | available | 5 | 2020-07|
----------------------------------------
|1 | available | 8 | 2020-08|
----------------------------------------
|1 | available | 8 | 2020-09|
----------------------------------------
|1 | limited | 8 | 2020-10|
----------------------------------------
|1 | limited | 8 | 2020-11|
----------------------------------------
|1 | limited | 8 | 2020-12|
----------------------------------------
|2 | limited | 1 | 2020-09|
----------------------------------------
|2 | limited | 3 | 2020-10|
----------------------------------------
|2 | limited | 3 | 2020-11|
----------------------------------------
|2 | limited | 3 | 2020-12|
----------------------------------------
请注意，对于product_id=1，有一个从2019-10年的旧记录，它被传播到2020-08年，然后被修剪，而对于product_id=2，在2020-09年之前没有记录，因此2020-07、2020-08月份没有被填充(因为在2020-0年之前不存在产品(。
由于数据帧由数百万条记录组成；暴力"；使用for循环并检查每个productid的解决方案相当缓慢。似乎应该可以使用窗口函数来解决这个问题，方法是在next_month创建另一列，然后根据该列填补空白，但我不知道如何实现这一点。
关于@jxc注释，我已经为这个用例准备好了答案。
以下是代码片段。
-
导入spark SQL函数
from pyspark.sql import functions as F, Window -
准备样本数据
simpleData = ((1,"Available",5,"2020-07"),
(1,"Available",8,"2020-08"),
(1,"Limited",8,"2020-12"),
(2,"Limited",1,"2020-09"),
(2,"Limited",3,"2020-12")
)
columns= ["product_id", "status", "price", "month"]
创建样本数据的数据帧
df = spark.createDataFrame(data = simpleData, schema = columns)在数据帧中添加日期列以获得正确格式的日期
df0 = df.withColumn("date",F.to_date('month','yyyy-MM'))
df0.show()
+----------+---------+-----+-------+----------+
|product_id| status|price| month| date|
+----------+---------+-----+-------+----------+
| 1|Available| 5|2020-07|2020-07-01|
| 1|Available| 8|2020-08|2020-08-01|
| 1| Limited| 8|2020-12|2020-12-01|
| 2| Limited| 1|2020-09|2020-09-01|
| 2| Limited| 3|2020-12|2020-12-01|
+----------+---------+-----+-------+----------+
- 创建WinSpec w1并使用Window聚合函数lead查找(w1(之后的下一个日期，将其转换为前几个月以设置日期序列：
w1 = Window.partitionBy('product_id').orderBy('date')
df1 = df0.withColumn('end_date',F.coalesce(F.add_months(F.lead('date').over(w1),-1),'date'))
df1.show()
+----------+---------+-----+-------+----------+----------+
|product_id| status|price| month| date| end_date|
+----------+---------+-----+-------+----------+----------+
| 1|Available| 5|2020-07|2020-07-01|2020-07-01|
| 1|Available| 8|2020-08|2020-08-01|2020-11-01|
| 1| Limited| 8|2020-12|2020-12-01|2020-12-01|
| 2| Limited| 1|2020-09|2020-09-01|2020-11-01|
| 2| Limited| 3|2020-12|2020-12-01|2020-12-01|
+----------+---------+-----+-------+----------+----------+
- 使用months_abetween(end_date，date(来计算两个日期之间的月份数，并使用transform函数迭代序列(0，#months(，创建一个named_struct，date=add_months(date，i(，price=IF(i=0，price，price(，使用inline_eouter来分解结构数组
df2 = df1.selectExpr("product_id", "status", inline_outer( transform( sequence(0,int(months_between(end_date, date)),1), i -> (add_months(date,i) as date, IF(i=0,price,price) as price) ) ) )
df2.show()
+----------+---------+----------+-----+
|product_id| status| date|price|
+----------+---------+----------+-----+
| 1|Available|2020-07-01| 5|
| 1|Available|2020-08-01| 8|
| 1|Available|2020-09-01| 8|
| 1|Available|2020-10-01| 8|
| 1|Available|2020-11-01| 8|
| 1| Limited|2020-12-01| 8|
| 2| Limited|2020-09-01| 1|
| 2| Limited|2020-10-01| 1|
| 2| Limited|2020-11-01| 1|
| 2| Limited|2020-12-01| 3|
+----------+---------+----------+-----+
- 对
product_id上的数据帧进行分区，并在df3中添加一个秩列，以获得每行的行号。然后，为每个product_id存储具有新列max_rank的rank列值的最大值，并将max_rank存储到df4
w2 = Window.partitionBy('product_id').orderBy('date')
df3 = df2.withColumn('rank',F.row_number().over(w2))
Schema: DataFrame[product_id: bigint, status: string, date: date, price: bigint, rank: int]
df3.show()
+----------+---------+----------+-----+----+
|product_id| status| date|price|rank|
+----------+---------+----------+-----+----+
| 1|Available|2020-07-01| 5| 1|
| 1|Available|2020-08-01| 8| 2|
| 1|Available|2020-09-01| 8| 3|
| 1|Available|2020-10-01| 8| 4|
| 1|Available|2020-11-01| 8| 5|
| 1| Limited|2020-12-01| 8| 6|
| 2| Limited|2020-09-01| 1| 1|
| 2| Limited|2020-10-01| 1| 2|
| 2| Limited|2020-11-01| 1| 3|
| 2| Limited|2020-12-01| 3| 4|
+----------+---------+----------+-----+----+
df4 = df3.groupBy("product_id").agg(F.max('rank').alias('max_rank'))
Schema: DataFrame[product_id: bigint, max_rank: int]
df4.show()
+----------+--------+
|product_id|max_rank|
+----------+--------+
| 1| 6|
| 2| 4|
+----------+--------+
- 在
product_id上加入df3和df4数据帧得到max_rank
df5 = df3.join(df4,df3.product_id == df4.product_id,"inner")
.select(df3.product_id,df3.status,df3.date,df3.price,df3.rank,df4.max_rank)
Schema: DataFrame[product_id: bigint, status: string, date: date, price: bigint, rank: int, max_rank: int]
df5.show()
+----------+---------+----------+-----+----+--------+
|product_id| status| date|price|rank|max_rank|
+----------+---------+----------+-----+----+--------+
| 1|Available|2020-07-01| 5| 1| 6|
| 1|Available|2020-08-01| 8| 2| 6|
| 1|Available|2020-09-01| 8| 3| 6|
| 1|Available|2020-10-01| 8| 4| 6|
| 1|Available|2020-11-01| 8| 5| 6|
| 1| Limited|2020-12-01| 8| 6| 6|
| 2| Limited|2020-09-01| 1| 1| 4|
| 2| Limited|2020-10-01| 1| 2| 4|
| 2| Limited|2020-11-01| 1| 3| 4|
| 2| Limited|2020-12-01| 3| 4| 4|
+----------+---------+----------+-----+----+--------+
- 然后使用
between函数最终过滤df5数据帧，以获得最近6个月的数据
FinalResultDF = df5.filter(F.col('rank')
.between(F.when((F.col('max_rank') > 5),(F.col('max_rank')-6)).otherwise(0),F.col('max_rank')))
.select(df5.product_id,df5.status,df5.date,df5.price)
FinalResultDF.show(truncate=False)
+----------+---------+----------+-----+
|product_id|status |date |price|
+----------+---------+----------+-----+
|1 |Available|2020-07-01|5 |
|1 |Available|2020-08-01|8 |
|1 |Available|2020-09-01|8 |
|1 |Available|2020-10-01|8 |
|1 |Available|2020-11-01|8 |
|1 |Limited |2020-12-01|8 |
|2 |Limited |2020-09-01|1 |
|2 |Limited |2020-10-01|1 |
|2 |Limited |2020-11-01|1 |
|2 |Limited |2020-12-01|3 |
+----------+---------+----------+-----+
使用spark-sql:
给定输入数据帧：
val df = spark.sql(""" with t1 (
select 1 c1, 'available' c2, 5 c3, '2019-10' c4 union all
select 1 c1, 'available' c2, 8 c3, '2020-08' c4 union all
select 1 c1, 'limited' c2, 8 c3, '2020-10' c4 union all
select 2 c1, 'limited' c2, 1 c3, '2020-09' c4 union all
select 2 c1, 'limited' c2, 3 c3, '2020-10' c4
) select c1 product_id, c2 status , c3 price, c4 month from t1
""")
df.createOrReplaceTempView("df")
df.show(false)
+----------+---------+-----+-------+
|product_id|status |price|month |
+----------+---------+-----+-------+
|1 |available|5 |2019-10|
|1 |available|8 |2020-08|
|1 |limited |8 |2020-10|
|2 |limited |1 |2020-09|
|2 |limited |3 |2020-10|
+----------+---------+-----+-------+
过滤日期窗口，即2020-07至2020-12的6个月，并将其存储在df1 中
val df1 = spark.sql("""
select * from df where month > '2020-07' and month < '2020-12'
""")
df1.createOrReplaceTempView("df1")
df1.show(false)
+----------+---------+-----+-------+
|product_id|status |price|month |
+----------+---------+-----+-------+
|1 |available|8 |2020-08|
|1 |limited |8 |2020-10|
|2 |limited |1 |2020-09|
|2 |limited |3 |2020-10|
+----------+---------+-----+-------+
下限-当月份<='时获得最大值2020-07'。将月份覆盖为"2020-07">
val df2 = spark.sql("""
select product_id, status, price, '2020-07' month from df where (product_id,month) in
( select product_id, max(month) from df where month <= '2020-07' group by 1 )
""")
df2.createOrReplaceTempView("df2")
df2.show(false)
+----------+---------+-----+-------+
|product_id|status |price|month |
+----------+---------+-----+-------+
|1 |available|5 |2020-07|
+----------+---------+-----+-------+
上限-使用<='获取最大值2020-12’。将月份覆盖为"2020-12">
val df3 = spark.sql("""
select product_id, status, price, '2020-12' month from df where (product_id, month) in
( select product_id, max(month) from df where month <= '2020-12' group by 1 )
""")
df3.createOrReplaceTempView("df3")
df3.show(false)
+----------+-------+-----+-------+
|product_id|status |price|month |
+----------+-------+-----+-------+
|1 |limited|8 |2020-12|
|2 |limited|3 |2020-12|
+----------+-------+-----+-------+
现在将所有3合并并存储在df4 中
val df4 = spark.sql("""
select product_id, status, price, month from df1 union all
select product_id, status, price, month from df2 union all
select product_id, status, price, month from df3
order by product_id, month
""")
df4.createOrReplaceTempView("df4")
df4.show(false)
+----------+---------+-----+-------+
|product_id|status |price|month |
+----------+---------+-----+-------+
|1 |available|5 |2020-07|
|1 |available|8 |2020-08|
|1 |limited |8 |2020-10|
|1 |limited |8 |2020-12|
|2 |limited |1 |2020-09|
|2 |limited |3 |2020-10|
|2 |limited |3 |2020-12|
+----------+---------+-----+-------+
结果：使用序列(date1，date2，interval 1 month(为缺少的月份生成日期数组。分解数组即可得到结果。
spark.sql("""
select product_id, status, price, month, explode(dt) res_month from
(
select t1.*,
case when months_between(lm||'-01',month||'-01')=1.0 then array(month||'-01')
when month='2020-12' then array(month||'-01')
else sequence(to_date(month||'-01'), add_months(to_date(lm||'-01'),-1), interval 1 month )
end dt
from (
select product_id, status, price, month,
lead(month) over(partition by product_id order by month) lm
from df4
) t1
) t2
order by product_id, res_month
""")
.show(false)
+----------+---------+-----+-------+----------+
|product_id|status |price|month |res_month |
+----------+---------+-----+-------+----------+
|1 |available|5 |2020-07|2020-07-01|
|1 |available|8 |2020-08|2020-08-01|
|1 |available|8 |2020-08|2020-09-01|
|1 |limited |8 |2020-10|2020-10-01|
|1 |limited |8 |2020-10|2020-11-01|
|1 |limited |8 |2020-12|2020-12-01|
|2 |limited |1 |2020-09|2020-09-01|
|2 |limited |3 |2020-10|2020-10-01|
|2 |limited |3 |2020-10|2020-11-01|
|2 |limited |3 |2020-12|2020-12-01|
+----------+---------+-----+-------+----------+