Cumulative product in Spark?

问题:

I try to implement a Cumululative product in Spark scala but I really don't know how to it. I have the following dataframe:

Input data:
+--+--+--------+----+
|A |B | date   | val|
+--+--+--------+----+
|rr|gg|20171103| 2  |
|hh|jj|20171103| 3  |
|rr|gg|20171104| 4  |
|hh|jj|20171104| 5  |
|rr|gg|20171105| 6  |
|hh|jj|20171105| 7  |
+-------+------+----+

And I would like to have the following output

Output data:
+--+--+--------+-----+
|A |B | date   | val |
+--+--+--------+-----+
|rr|gg|20171105| 48  | // 2 * 4 * 6
|hh|jj|20171105| 105 | // 3 * 5 * 7
+-------+------+-----+

If you have any idea about how to do it, it would be really helpful :)

Thank a lot

回答1:

As long as the number are strictly positive (0 can be handled as well, if present, using coalesce) as in your example, the simplest solution is to compute the sum of logarithms and take the exponential:

import org.apache.spark.sql.functions.{exp, log, max, sum}

val df = Seq(
  ("rr", "gg", "20171103", 2), ("hh", "jj", "20171103", 3), 
  ("rr", "gg", "20171104", 4), ("hh", "jj", "20171104", 5), 
  ("rr", "gg", "20171105", 6), ("hh", "jj", "20171105", 7)
).toDF("A", "B", "date", "val")

val result = df
  .groupBy("A", "B")
  .agg(
    max($"date").as("date"), 
    exp(sum(log($"val"))).as("val"))

Since this uses FP arithmetic the result won't be exact:

result.show

+---+---+--------+------------------+
|  A|  B|    date|               val|
+---+---+--------+------------------+
| hh| jj|20171105|104.99999999999997|
| rr| gg|20171105|47.999999999999986|
+---+---+--------+------------------+

but after rounding should good enough for majority of applications.

result.withColumn("val", round($"val")).show

+---+---+--------+-----+
|  A|  B|    date|  val|
+---+---+--------+-----+
| hh| jj|20171105|105.0|
| rr| gg|20171105| 48.0|
+---+---+--------+-----+

If that's not enough you can define an UserDefinedAggregateFunction or Aggregator (How to define and use a User-Defined Aggregate Function in Spark SQL?) or use functional API with reduceGroups:

import scala.math.Ordering

case class Record(A: String, B: String, date: String, value: Long)

df.withColumnRenamed("val", "value").as[Record]
  .groupByKey(x => (x.A, x.B))
  .reduceGroups((x, y) => x.copy(
    date = Ordering[String].max(x.date, y.date),
    value = x.value * y.value))
  .toDF("key", "value")
  .select($"value.*")
  .show

+---+---+--------+-----+
|  A|  B|    date|value|
+---+---+--------+-----+
| hh| jj|20171105|  105|
| rr| gg|20171105|   48|
+---+---+--------+-----+

回答2:

You can solve this using either collect_list+UDF or an UDAF. UDAF may be more efficient, but harder to implement due to the local aggregation.

If you have a dataframe like this :

+---+---+
|key|val|
+---+---+
|  a|  1|
|  a|  2|
|  a|  3|
|  b|  4|
|  b|  5|
+---+---+

You can invoke an UDF :

val prod = udf((vals:Seq[Int]) => vals.reduce(_ * _))

df
  .groupBy($"key")
  .agg(prod(collect_list($"val")).as("val"))
  .show()

+---+---+
|key|val|
+---+---+
|  b| 20|
|  a|  6|
+---+---+