将递归SQL转换转换为Spark

问题描述

我们正在获取具有以下字段的订单数据(仅显示相关字段)

enter image description here

  1. original_orderid为NULL的订单可以被视为“父订单”
  2. 其中一些父订单可能具有子订单,而子订单的孩子的original_orderid映射到父母的订单ID。
  3. 子订单可以生成一个子订单,如图所示,带有颜色编码

与原始文本相同的数据:

+-----------+----------------+---------------------+----------+
|orderid    |original_orderid|ttime                |price     |
+-----------+----------------+---------------------+----------+
|988782828  |0               |2020-09-0406:00:09.09|3444.0    |
|37377373374|0               |2020-09-0408:41:09.09|26262.0   |
|23222223378|37377373374     |2020-09-0409:02:55.55|33434.0   |
|2111111    |0               |2020-09-0409:05:55.55|44334.0   |
|2422244422 |0               |2020-09-0409:07:14.14|343434.0  |
|66666663388|23222223378     |2020-09-0409:10:14.14|1282.0    |
|44444443391|66666663388     |2020-09-0409:11:34.34|27272.6363|
|22222393392|44444443391     |2020-09-0409:13:38.38|333.0     |
|77777393397|22222393392     |2020-09-0409:14:31.31|3422.0    |
|55656563397|77777393397     |2020-09-0409:16:58.58|27272.0   |
+-----------+----------------+---------------------+----------+

作为转换,我们需要将所有子级映射到其原始父级(original_orderid为NULL),并获取订单可能具有的级别数。 预期结果将是:

enter image description here

这是从sqlserver到spark迁移工作的一部分。在sql server中,这是通过递归访问父视图的视图实现的。

我们可以使用代码在Spark中尝试这种转换:

val df = spark.read(raw_data_file)
val parent = df.filter(col(original_orderid).isNull)
                .select(col(orderid).as("orderid"),col(order_id).as("parent_orderid")

val children = df.filter(col(original_orderid).isNotNull).sort(col(ttime))


var prentCollection = //Collect parent df in collection 
val childrenCollection = //Collect child df in collection 

//Traverse through the sorted childrenCollection
for (child <- childrenCollection) ={
    if child.original_orderid in parentCollection.orderid.alias(parent)
    
    insert into parentCollection - child.orderid as orderid,parent.parent_orderid as parent_orderid,child.ttime as ttime,child.price as price
}

解决方案需要在驱动程序中收集所有数据,因此无法分发并且不适合大型数据集。

您能否建议我使用其他任何方法使它适用于较大的数据集,或者对上述现有方法进行任何改进。

解决方法

您可以递归地加入并累积数组中的父母。这是一个使用Spark v2.1的快速原型

val addToArray = udf((seq : Seq[String],item: String) => seq :+ item)
//v2.4.0 use array_union 
val concatArray = udf((seq1 : Seq[String],seq2 : Seq[String]) => seq1 ++ seq2)
//v2.4.0 use element_at and size 
val lastInArray = udf((seq: Seq[String]) => seq.lastOption.getOrElse(null))
//v2.4.0 and up use slice
val dropLastInArray = udf((seq: Seq[String]) => seq.dropRight(1))

val raw="""|988782828  |0               |2020-09-0406:00:09.09|3444.0    |
|37377373374|0               |2020-09-0408:41:09.09|26262.0   |
|23222223378|37377373374     |2020-09-0409:02:55.55|33434.0   |
|2111111    |0               |2020-09-0409:05:55.55|44334.0   |
|2422244422 |0               |2020-09-0409:07:14.14|343434.0  |
|66666663388|23222223378     |2020-09-0409:10:14.14|1282.0    |
|44444443391|66666663388     |2020-09-0409:11:34.34|27272.6363|
|22222393392|44444443391     |2020-09-0409:13:38.38|333.0     |
|77777393397|22222393392     |2020-09-0409:14:31.31|3422.0    |
|55656563397|77777393397     |2020-09-0409:16:58.58|27272.0   |"""

val df= raw.substring(1).split("\\n").map(_.split("\\|").map(_.trim)).map(r=> (r(0),r(1),r(2),r(3))).toSeq.toDF        ("orderId","parentId","ttime","price").withColumn("parents",array(col("parentId"))) 

def selfJoin(df :DataFrame) : DataFrame = {
      if (df.filter(lastInArray(col("parents")) =!= lit("0")).count > 0)
        selfJoin(df.join(df.select(col("orderId").as("id"),col("parents").as("grandParents")),lastInArray(col("parents"))  === col("id"),"left").withColumn("parents",when(lastInArray(col("parents")) =!= lit("0"),concatArray(col("parents"),col("grandParents"))).otherwise(col("parents"))).drop("grandParents").drop("id"))
      else
      df
     
}

selfJoin(df).withColumn("level",size(col("parents"))).withColumn("top parent",lastInArray(dropLastInArray(col("parents")))).show