Spark Advanced Window with dynamic last

Question

Problem: Given a time series data which is a clickstream of user activity is stored in hive, ask is to enrich the data with session id using spark.

Session Definition

Answer 1

Not a straight forward problem to solve, but here's one approach:

1. Use Window lag timestamp difference to identify sessions (with 0 = start of a session) per user for rule #1
2. Group the dataset to assemble the timestamp diff list per user
3. Process via a UDF the timestamp diff list to identify sessions for rule #2 and create all session ids per user
4. Expand the grouped dataset via Spark's explode

Sample code below:

import org.apache.spark.sql.functions._
import org.apache.spark.sql.expressions.Window
import spark.implicits._

val userActivity = Seq(
  ("2018-01-01 11:00:00", "u1"),
  ("2018-01-01 12:10:00", "u1"),
  ("2018-01-01 13:00:00", "u1"),
  ("2018-01-01 13:50:00", "u1"),
  ("2018-01-01 14:40:00", "u1"),
  ("2018-01-01 15:30:00", "u1"),
  ("2018-01-01 16:20:00", "u1"),
  ("2018-01-01 16:50:00", "u1"),
  ("2018-01-01 11:00:00", "u2"),
  ("2018-01-02 11:00:00", "u2")
).toDF("click_time", "user_id")

def clickSessList(tmo: Long) = udf{ (uid: String, clickList: Seq[String], tsList: Seq[Long]) =>
  def sid(n: Long) = s"$uid-$n"

  val sessList = tsList.foldLeft( (List[String](), 0L, 0L) ){ case ((ls, j, k), i) =>
    if (i == 0 || j + i >= tmo) (sid(k + 1) :: ls, 0L, k + 1) else
       (sid(k) :: ls, j + i, k)
  }._1.reverse

  clickList zip sessList
}

Note that the accumulator for foldLeft in the UDF is a Tuple of (ls, j, k), where:

• ls is the list of formatted session ids to be returned
• j and k are for carrying over the conditionally changing timestamp value and session id number, respectively, to the next iteration

Step 1:

val tmo1: Long = 60 * 60
val tmo2: Long = 2 * 60 * 60

val win1 = Window.partitionBy("user_id").orderBy("click_time")

val df1 = userActivity.
  withColumn("ts_diff", unix_timestamp($"click_time") - unix_timestamp(
    lag($"click_time", 1).over(win1))
  ).
  withColumn("ts_diff", when(row_number.over(win1) === 1 || $"ts_diff" >= tmo1, 0L).
    otherwise($"ts_diff")
  )

df1.show
// +-------------------+-------+-------+
// |         click_time|user_id|ts_diff|
// +-------------------+-------+-------+
// |2018-01-01 11:00:00|     u1|      0|
// |2018-01-01 12:10:00|     u1|      0|
// |2018-01-01 13:00:00|     u1|   3000|
// |2018-01-01 13:50:00|     u1|   3000|
// |2018-01-01 14:40:00|     u1|   3000|
// |2018-01-01 15:30:00|     u1|   3000|
// |2018-01-01 16:20:00|     u1|   3000|
// |2018-01-01 16:50:00|     u1|   1800|
// |2018-01-01 11:00:00|     u2|      0|
// |2018-01-02 11:00:00|     u2|      0|
// +-------------------+-------+-------+

Steps 2-4:

val df2 = df1.
  groupBy("user_id").agg(
    collect_list($"click_time").as("click_list"), collect_list($"ts_diff").as("ts_list")
  ).
  withColumn("click_sess_id",
    explode(clickSessList(tmo2)($"user_id", $"click_list", $"ts_list"))
  ).
  select($"user_id", $"click_sess_id._1".as("click_time"), $"click_sess_id._2".as("sess_id"))

df2.show
// +-------+-------------------+-------+
// |user_id|click_time         |sess_id|
// +-------+-------------------+-------+
// |u1     |2018-01-01 11:00:00|u1-1   |
// |u1     |2018-01-01 12:10:00|u1-2   |
// |u1     |2018-01-01 13:00:00|u1-2   |
// |u1     |2018-01-01 13:50:00|u1-2   |
// |u1     |2018-01-01 14:40:00|u1-3   |
// |u1     |2018-01-01 15:30:00|u1-3   |
// |u1     |2018-01-01 16:20:00|u1-3   |
// |u1     |2018-01-01 16:50:00|u1-4   |
// |u2     |2018-01-01 11:00:00|u2-1   |
// |u2     |2018-01-02 11:00:00|u2-2   |
// +-------+-------------------+-------+

Also note that click_time is "passed thru" in steps 2-4 so as to be included in the final dataset.