在Spark中计算UDF的调用

Question

Using Spark 1.6.1 I want to call the number of times a UDF is called. I want to do this because I have a very expensive UDF (~1sec per call) and I suspect the UDF being called more often than the number of records in my dataframe, making my spark job slower than necessary .

Although I could not reproduce this situation, I came up with a simple example showing that the number of calls to the UDF seems to be different (here: less) than the number of rows, how can that be?

import org.apache.spark.sql.SQLContext
import org.apache.spark.{SparkConf, SparkContext}
import org.apache.spark.sql.functions.udf

object Demo extends App {
  val conf = new SparkConf().setMaster("local[4]").setAppName("Demo")
  val sc = new SparkContext(conf)
  sc.setLogLevel("WARN")
  val sqlContext = new SQLContext(sc)
  import sqlContext.implicits._


  val callCounter = sc.accumulator(0)

  val df= sc.parallelize(1 to 10000,numSlices = 100).toDF("value")

  println(df.count) //  gives 10000

  val myudf = udf((d:Int) => {callCounter.add(1);d})

  val res = df.withColumn("result",myudf($"value")).cache

  println(res.select($"result").collect().size) // gives 10000
  println(callCounter.value) // gives 9941

}

If using an accumulator is not the right way to call the counts of the UDF, how else could I do it?

Note: In my actual Spark-Job, get a call-count which is about 1.7 times higher than the actual number of records.

Answer 1

Spark applications should define a main() method instead of extending scala.App. Subclasses of scala.App may not work correctly.

import org.apache.spark.sql.SQLContext
import org.apache.spark.{SparkConf, SparkContext}
import org.apache.spark.sql.functions.udf

object Demo extends App {
    def main(args: Array[String]): Unit = {
         val conf = new SparkConf().setAppName("Simple Application").setMaster("local[4]")
         val sc = new SparkContext(conf)
         // [...]
    }   
}

This should solve your problem.