How to aggregate values into collection after groupBy?

@zero323's answer is pretty much complete, but Spark gives us even more flexibility. How about the following solution?

import org.apache.spark.sql.functions._
inventory
  .select($"*", explode($"trackingIds") as "tracking_id")
  .select($"*", explode($"emailIds") as "email_id")
  .groupBy("visitorId")
  .agg(
    collect_list("tracking_id") as "trackingIds",
    collect_list("email_id") as "emailIds")

That however leaves out all empty collections (so there's some room for improvement :))

You can use User defined aggregated functions.

1) create a custom UDAF using the scala class called customAggregation.

package com.package.name

import org.apache.spark.sql.Row
import org.apache.spark.sql.expressions.{MutableAggregationBuffer, UserDefinedAggregateFunction}
import org.apache.spark.sql.types._
import scala.collection.JavaConverters._

class CustomAggregation() extends UserDefinedAggregateFunction {

// Input Data Type Schema
def inputSchema: StructType = StructType(Array(StructField("col5", ArrayType(StringType))))

// Intermediate Schema
def bufferSchema = StructType(Array(
StructField("col5_collapsed",  ArrayType(StringType))))

// Returned Data Type .
def dataType: DataType = ArrayType(StringType)

// Self-explaining
def deterministic = true

// This function is called whenever key changes
def initialize(buffer: MutableAggregationBuffer) = {
buffer(0) = Array.empty[String] // initialize array
}

// Iterate over each entry of a group
def update(buffer: MutableAggregationBuffer, input: Row) = {
buffer(0) =
  if(!input.isNullAt(0))
    buffer.getList[String](0).toArray ++ input.getList[String](0).toArray
  else
    buffer.getList[String](0).toArray
}

  // Merge two partial aggregates
 def merge(buffer1: MutableAggregationBuffer, buffer2: Row) = {
 buffer1(0) = buffer1.getList[String](0).toArray ++ buffer2.getList[String](0).toArray
}

 // Called after all the entries are exhausted.
 def evaluate(buffer: Row) = {
  buffer.getList[String](0).asScala.toList.distinct
 }
}

2) Then use the UDAF in your code as

//define UDAF
val CustomAggregation = new CustomAggregation()
DataFrame
    .groupBy(col1,col2,col3)
    .agg(CustomAggregation(DataFrame(col5))).show()

Spark >= 2.4

You can replace flatten udf with built-in flatten function

import org.apache.spark.sql.functions.flatten

leaving the rest as-is.

Spark >= 2.0, < 2.4

It is possible but quite expensive. Using data you've provided:

case class Record(
    visitorId: String, trackingIds: Array[String], emailIds: Array[String])

val df = Seq(
  Record("a158", Array("666b"), Array("12")),
  Record("7g21", Array("c0b5"), Array("45")),
  Record("7g21", Array("c0b4"), Array("87")),
  Record("a158", Array("666b",  "777c"), Array.empty[String])).toDF

and a helper function:

import org.apache.spark.sql.functions.udf

val flatten = udf((xs: Seq[Seq[String]]) => xs.flatten)

we can fill the blanks with placeholders:

import org.apache.spark.sql.functions.{array, lit, when}

val dfWithPlaceholders = df.withColumn(
  "emailIds", 
  when(size($"emailIds") === 0, array(lit(""))).otherwise($"emailIds"))

collect_lists and flatten:

import org.apache.spark.sql.functions.{array, collect_list}

val emailIds = flatten(collect_list($"emailIds")).alias("emailIds")
val trackingIds = flatten(collect_list($"trackingIds")).alias("trackingIds")

df
  .groupBy($"visitorId")
  .agg(trackingIds, emailIds)

// +---------+------------------+--------+
// |visitorId|       trackingIds|emailIds|
// +---------+------------------+--------+
// |     a158|[666b, 666b, 777c]|  [12, ]|
// |     7g21|      [c0b5, c0b4]|[45, 87]|
// +---------+------------------+--------+

With statically typed Dataset:

df.as[Record]
  .groupByKey(_.visitorId)
  .mapGroups { case (key, vs) => 
    vs.map(v => (v.trackingIds, v.emailIds)).toArray.unzip match {
      case (trackingIds, emailIds) => 
        Record(key, trackingIds.flatten, emailIds.flatten)
  }}

// +---------+------------------+--------+
// |visitorId|       trackingIds|emailIds|
// +---------+------------------+--------+
// |     a158|[666b, 666b, 777c]|  [12, ]|
// |     7g21|      [c0b5, c0b4]|[45, 87]|
// +---------+------------------+--------+

Spark 1.x

You can convert to RDD and group

import org.apache.spark.sql.Row

dfWithPlaceholders.rdd
  .map {
     case Row(id: String, 
       trcks: Seq[String @ unchecked],
       emails: Seq[String @ unchecked]) => (id, (trcks, emails))
  }
  .groupByKey
  .map {case (key, vs) => vs.toArray.unzip match {
    case (trackingIds, emailIds) => 
      Record(key, trackingIds.flatten, emailIds.flatten)
  }}
  .toDF

// +---------+------------------+--------+
// |visitorId|       trackingIds|emailIds|
// +---------+------------------+--------+
// |     7g21|      [c0b5, c0b4]|[45, 87]|
// |     a158|[666b, 666b, 777c]|  [12, ]|
// +---------+------------------+--------+