How to use “cube” only for specific fields on Spark dataframe?

Question

I'm using Spark 1.6.1, and I have such a dataframe.

+-------------+-----------+-----------------+-------+-------+-------+----------+-------+-------+-------+-------+
|     scene_id|  action_id|       classifier|os_name|country|app_ver|   p0value|p1value|p2value|p3value|p4value|
+-------------+-----------+-----------------+-------+-------+-------+----------+-------+-------+-------+-------+
|    test_home|scene_enter|        test_home|android|     KR|  5.6.3|__OTHERS__|  false|   test|   test|   test|
......

And I want to get dataframe like as following by using cube operation.

(Grouped by all fields, but only "os_name", "country", "app_ver" fields are cubed)

+-------------+-----------+-----------------+-------+-------+-------+----------+-------+-------+-------+-------+---+
|     scene_id|  action_id|       classifier|os_name|country|app_ver|   p0value|p1value|p2value|p3value|p4value|cnt|
+-------------+-----------+-----------------+-------+-------+-------+----------+-------+-------+-------+-------+---+
|    test_home|scene_enter|        test_home|android|     KR|  5.6.3|__OTHERS__|  false|   test|   test|   test|  9|
|    test_home|scene_enter|        test_home|   null|     KR|  5.6.3|__OTHERS__|  false|   test|   test|   test| 35|
|    test_home|scene_enter|        test_home|android|   null|  5.6.3|__OTHERS__|  false|   test|   test|   test| 98|
|    test_home|scene_enter|        test_home|android|     KR|   null|__OTHERS__|  false|   test|   test|   test|101|
|    test_home|scene_enter|        test_home|   null|   null|  5.6.3|__OTHERS__|  false|   test|   test|   test|301|
|    test_home|scene_enter|        test_home|   null|     KR|   null|__OTHERS__|  false|   test|   test|   test|225|
|    test_home|scene_enter|        test_home|android|   null|   null|__OTHERS__|  false|   test|   test|   test|312|
|    test_home|scene_enter|        test_home|   null|   null|   null|__OTHERS__|  false|   test|   test|   test|521|
......

I have tried like below, but it seems to be slow and ugly..

var cubed = df
  .cube($"scene_id", $"action_id", $"classifier", $"country", $"os_name", $"app_ver", $"p0value", $"p1value", $"p2value", $"p3value", $"p4value")
  .count
  .where("scene_id IS NOT NULL AND action_id IS NOT NULL AND classifier IS NOT NULL AND p0value IS NOT NULL AND p1value IS NOT NULL AND p2value IS NOT NULL AND p3value IS NOT NULL AND p4value IS NOT NULL")

Any better solutions?

Answer 1

I believe you cannot avoid the problem completely but there is a simple trick you can reduce its scale. The idea is to replace all columns, which shouldn't be marginalized, with a single placeholder.

For example if you have a DataFrame:

val df = Seq((1, 2, 3, 4, 5, 6)).toDF("a", "b", "c", "d", "e", "f")

and you're interested in cube marginalized by d and e and grouped by a..c you can define the substitute for a..c as:

import org.apache.spark.sql.functions.struct
import sparkSql.implicits._

// alias here may not work in Spark 1.6
val rest = struct(Seq($"a", $"b", $"c"): _*).alias("rest")

and cube:

val cubed =  Seq($"d", $"e")

// If there is a problem with aliasing rest it can done here.
val tmp = df.cube(rest.alias("rest") +: cubed: _*).count

Quick filter and select should handle the rest:

tmp.where($"rest".isNotNull).select($"rest.*" +: cubed :+ $"count": _*)

with result like:

+---+---+---+----+----+-----+
|  a|  b|  c|   d|   e|count|
+---+---+---+----+----+-----+
|  1|  2|  3|null|   5|    1|
|  1|  2|  3|null|null|    1|
|  1|  2|  3|   4|   5|    1|
|  1|  2|  3|   4|null|    1|
+---+---+---+----+----+-----+