Question

我有一个时间序列数据集，该数据集按ID划分，并按时间戳排序。样本：

  ID     Timestamp   Feature

 "XSC"   1986-05-21  44.7530
 "XSC"   1986-05-22  44.7530
 "XSC"   1986-05-23  23.5678

 "TM"    1982-03-08  22.2734
 "TM"    1982-03-09  22.1941
 "TM"    1982-03-10  22.0847
 "TM"    1982-03-11  22.1741
 "TM"    1982-03-12  22.1840
 "TM"    1982-03-15  22.1344

我有一些需要计算的自定义逻辑，应该在每个分区的每个窗口中完成。我知道Spark为窗口功能提供了丰富的支持，我正试图将其用于此目的。

我的逻辑要求当前窗口/分区中的元素总数为标量。我需要执行一些特定的计算（基本上是一个for循环，直到该计数）。

我尝试通过

添加一个count列

val window = Window.partitionBy("id").orderBy("timestamp") 
frame = frame.withColumn("my_cnt", count(column).over(window))

我需要做类似的事情：

var i = 1
var y = col("Feature")
var result = y
while (i < /* total number of records within each partition goes here */) {
    result = result + lit(1) * lag(y, i).over(window) + /* complex computation */
    i = i + 1
}
dataFrame.withColumn("Computed_Value", result)

如何获取每个分区中的记录总数作为标量值？我还添加了该计数“ my_cnt”值，这增加了分区的总值，但在我的情况下似乎无法使用它。

Answer 1

Spark的collect_list功能可让您将窗口值汇总为列表。此列表可以传递给udf进行一些复杂的计算

因此，如果您有来源

val data = List(
  ("XSC", "1986-05-21", 44.7530),
  ("XSC", "1986-05-22", 44.7530),
  ("XSC", "1986-05-23", 23.5678),
  ("TM", "1982-03-08", 22.2734),
  ("TM", "1982-03-09", 22.1941),
  ("TM", "1982-03-10", 22.0847),
  ("TM", "1982-03-11", 22.1741),
  ("TM", "1982-03-12", 22.1840),
  ("TM", "1982-03-15", 22.1344),
).toDF("id", "timestamp", "feature")
  .withColumn("timestamp", to_date('timestamp))

还有一些复杂的函数，包装在记录中的UDF中（例如，表示为元组）

 val complexComputationUDF = udf((list: Seq[Row]) => {
  list
    .map(row => (row.getString(0), row.getDate(1).getTime, row.getDouble(2)))
    .sortBy(-_._2)
    .foldLeft(0.0) {
      case (acc, (id, timestamp, feature)) => acc + feature
    }
})

您可以定义一个将所有分区数据传递到每个记录的窗口，或者在有序窗口的情况下，将一个增量数据传递到每个记录

val windowAll = Window.partitionBy("id")
val windowRunning = Window.partitionBy("id").orderBy("timestamp")

并将它们放到一个新的数据集中，例如：

val newData = data
  // I assuming thatyou need id,timestamp & feature for the complex computattion. So I create a struct
  .withColumn("record", struct('id, 'timestamp, 'feature))
  // Collect all records in the partition as a list of tuples and pass them to the complexComupation
  .withColumn("computedValueAll",
     complexComupationUDF(collect_list('record).over(windowAll)))
  // Collect records in a time ordered windows in the partition as a list of tuples and pass them to the complexComupation
  .withColumn("computedValueRunning",
     complexComupationUDF(collect_list('record).over(windowRunning)))

这将导致类似以下内容：

+---+----------+-------+--------------------------+------------------+--------------------+
|id |timestamp |feature|record                    |computedValueAll  |computedValueRunning|
+---+----------+-------+--------------------------+------------------+--------------------+
|XSC|1986-05-21|44.753 |[XSC, 1986-05-21, 44.753] |113.07379999999999|44.753              |
|XSC|1986-05-22|44.753 |[XSC, 1986-05-22, 44.753] |113.07379999999999|89.506              |
|XSC|1986-05-23|23.5678|[XSC, 1986-05-23, 23.5678]|113.07379999999999|113.07379999999999  |
|TM |1982-03-08|22.2734|[TM, 1982-03-08, 22.2734] |133.0447          |22.2734             |
|TM |1982-03-09|22.1941|[TM, 1982-03-09, 22.1941] |133.0447          |44.4675             |
|TM |1982-03-10|22.0847|[TM, 1982-03-10, 22.0847] |133.0447          |66.5522             |
|TM |1982-03-11|22.1741|[TM, 1982-03-11, 22.1741] |133.0447          |88.7263             |
|TM |1982-03-12|22.184 |[TM, 1982-03-12, 22.184]  |133.0447          |110.91029999999999  |
|TM |1982-03-15|22.1344|[TM, 1982-03-15, 22.1344] |133.0447          |133.0447            |
+---+----------+-------+--------------------------+------------------+--------------------+

Spark Window自定义函数-获取分区记录的总数

1 个答案: