SparkSQL sum if on multiple conditions

问题:

I have a SparkSQL DataFrame like this:

name gender age isActive points
-------------------------------
 Bob      M  12     true    100
 Hal      M  16    false     80
 Pat      F  21     true     70
 Lin      F  17    false     40
 Zac      M  18     true     20
 Mei      F  19     true     10
 Sal      M  13    false     10

I have a few functions like so:

def isEligible(prog: String) (name: String, gender: String, age: Int, isActive: Boolean, points: Int): Boolean

that determines whether someone is eligible for a particular program. For Instance, the following function call would tell me whether Bob is eligible for Program1:

isEligible("Program1", "Bob", "M", 12, true, 100)

A person may be eligible for more than one program. I want to write a function that takes this DataFrame, and outputs a summary DataFrame like so:

prog1 prog2 prog3 prog4
-----------------------
    7     3     2     5

which shows the number of people who are eligible for each program. What is the best way to do this in Spark? I know I can use struct and agg functions, but I don't know how to incorporate my isEligible function into the SparkSQL query.

回答1:

Define a list of programs:

val progs = Seq("prog1", "prog2", "prog3", "prog4")

Define expressions

@transient val exprs = progs.map(p => {
  val f = udf(isEligible(p) _)
  sum(f(
    $"name", $"gender", $"age", $"isActive", $"points"
  ).cast("long")).alias(p)
})

df.select(exprs: _*)

You could also use strongly typed dataset:

import org.apache.spark.sql.Row

case class Record(name: String, gender: String, age: Int, 
                  isActive: Boolean, points: Int)

df.as[Record].flatMap {
   case Record(name, gender, age, isActive, points) => 
     progs.filter(p => isEligible(p)(name, gender, age, isActive, points))
}.groupBy().pivot("value", progs).count()