How to get Histogram of all columns in a large CSV / RDD[Array[double]] using Apache Spark Scala?

Question

I am trying to calculate Histogram of all columns from a CSV file using Spark Scala.

I found that DoubleRDDFunctions supporting Histogram. So I coded like following for getting histogram of all columns.

1. Get column count

2. Create RDD[double] of each column and calculate Histogram of each RDD using DoubleRDDFunctions

   var columnIndexArray = Array.tabulate(rdd.first().length) (_ * 1)
   
   val histogramData = columnIndexArray.map(columns => {
     rdd.map(lines => lines(columns)).histogram(6) 
   })
   

Is it a good way ? Can anyone suggest some better ways to tackle this ?

Thanks in advance.

Answer 1

Not exactly better but alternative way is to convert a RDD to a DataFrame and use histogram_numeric UDF.

Example data:

import scala.util.Random
import org.apache.spark.sql.types._
import org.apache.spark.sql.functions.{callUDF, lit, col}
import org.apache.spark.sql.Row
import org.apache.spark.sql.hive.HiveContext

val sqlContext = new HiveContext(sc)

Random.setSeed(1)

val ncol = 5

val rdd = sc.parallelize((1 to 1000).map(
  _ => Row.fromSeq(Array.fill(ncol)(Random.nextDouble))
))

val schema = StructType(
  (1 to ncol).map(i => StructField(s"x$i", DoubleType, false)))

val df = sqlContext.createDataFrame(rdd, schema)
df.registerTempTable("df")

Query:

val nBuckets = 3
val columns = df.columns.map(
  c => callUDF("histogram_numeric", col(c), lit(nBuckets)).alias(c))
val histograms = df.select(columns: _*)

histograms.printSchema

// root
//  |-- x1: array (nullable = true)
//  |    |-- element: struct (containsNull = true)
//  |    |    |-- x: double (nullable = true)
//  |    |    |-- y: double (nullable = true)
//  |-- x2: array (nullable = true)
//  |    |-- element: struct (containsNull = true)
//  |    |    |-- x: double (nullable = true)
//  |    |    |-- y: double (nullable = true)
//  |-- x3: array (nullable = true)
//  |    |-- element: struct (containsNull = true)
//  |    |    |-- x: double (nullable = true)
//  |    |    |-- y: double (nullable = true)
//  |-- x4: array (nullable = true)
//  |    |-- element: struct (containsNull = true)
//  |    |    |-- x: double (nullable = true)
//  |    |    |-- y: double (nullable = true)
//  |-- x5: array (nullable = true)
//  |    |-- element: struct (containsNull = true)
//  |    |    |-- x: double (nullable = true)
//  |    |    |-- y: double (nullable = true)

histograms.select($"x1").collect()

// Array([WrappedArray([0.16874313309969038,334.0],
//   [0.513382068667877,345.0], [0.8421388886903808,321.0])])

Answer 2

the (scala api) transformation, countByValue ought to do what you want

so for instance to generate histogram data for the first column in your RDD:

val histCol1 = RDD.map(record => record.col_1).countByValue()

in the expression above, record just refers to a data row in the RDD, an instance of a case class that has a field col_1

and so histCol1 will return a hash table (Scala Map) in which the keys are the unique values in column 1 (col_1) and the values are obviously the frequencies of each unique value