Apache Spark - Dealing with Sliding Windows on Temporal RDDs

Question

I\'ve been working quite a lot with Apache Spark the last few months but now I have received a pretty difficult task, to compute average/minimum/maximum etcetera on a sliding wi

Answer 1

If you convert to a DataFrame, this all gets a lot simpler -- you can just self-join the data back on itself and find the average. Say I have a series of data like this:

tsDF.show
date       amount
1970-01-01 10.0
1970-01-01 5.0
1970-01-01 7.0
1970-01-02 14.0
1970-01-02 13.9
1970-01-03 1.0
1970-01-03 5.0
1970-01-03 9.0
1970-01-04 9.0
1970-01-04 5.8
1970-01-04 2.8
1970-01-04 8.9
1970-01-05 8.1
1970-01-05 2.1
1970-01-05 2.78
1970-01-05 20.78

Which rolls up as:

tsDF.groupBy($"date").agg($"date", sum($"amount"), count($"date")).show
date       SUM(amount) COUNT(date)
1970-01-01 22.0        3
1970-01-02 27.9        2
1970-01-03 15.0        3
1970-01-04 26.5        4
1970-01-05 33.76       4

I then would need to create a UDF to shift the date for the join condition (note I am only using a 2 day window by using offset = -2):

def dateShift(myDate: java.sql.Date): java.sql.Date = {
  val offset = -2;
  val cal = Calendar.getInstance;
  cal.setTime(myDate);
  cal.add(Calendar.DATE, offset);
  new java.sql.Date(cal.getTime.getTime)
}
val udfDateShift = udf[java.sql.Date,java.sql.Date](dateShift)

And then I could easily find a 2-day rolling average like this:

val windowDF = tsDF.select($"date")
  .groupBy($"date")
  .agg($"date")
  .join(
    tsDF.select($"date" as "r_date", $"amount" as "r_amount"),
    $"r_date" > udfDateShift($"date") and $"r_date" <= $"date"
  )
  .groupBy($"date")
  .agg($"date",avg($"r_amount") as "2 day avg amount / record")

val windowDF.show
date       2 day avg amount / record
1970-01-01 7.333333333333333
1970-01-02 9.98
1970-01-03 8.58
1970-01-04 5.928571428571429
1970-01-05 7.5325

While this isn't exactly what you were trying to do, you see how you can use a DataFrame self-join to extract running averages from a data set. Hope you found this helpful.