stjordanis / spark-flow

Library for organizing batch processing pipelines in Apache Spark

spark-flow

This is a library for organizing batch processing pipelines in Apache Spark and handling automatic checkpointing of intermediate results. The core type is a DC (Distributed Collection) which is analogous to a Spark Dataset. DCs have an API encompassing RDDs, Datasets, and Dataframes.

A logical pipeline can be constructed entirely lazily without a Spark context. Pass in a Spark context to any DC to get the corresponding Dataset, RDD or Dataframe.

Building Locally

sbt publishLocal

Include in project

libraryDependencies += "com.bloomberg" %% "spark-flow" % "0.1.0"

Basic Example

import com.bloomberg.sparkflow

val numbers = sparkflow.parallelize(1 to 10)
val filtered = numbers.filter(_ < 3).checkpoint()
val doubled = filtered.map(_ * 2)

println(doubled.getRDD(sc).sum())

Combined Dataframe API

import com.bloomberg.sparkflow

val cars = sparkflow.read
  .format("csv")
  .option("header", "true")
  .load(testFile("cars.csv"))

val makeModel = cars.select("make", "model").checkpoint()

makeModel.getDF(sc).show()

Larger Example

object FilmsPipeline {

  class FilmMain(){

    val inputFilmRows = sparkflow.read.format("csv").option("header", "true")
      .load(testFile("Film_Locations_in_San_Francisco.csv"))
      .toDF(FilmsPipeline.columns:_*)
      .as[FilmsPipeline.InputFilmRow]

    val filmRows = inputFilmRows.keyBy(_.title)
      .groupByKey()
      .map(parseFilmRows)
      .checkpoint()

    val actorMovieCount = filmRows.flatMap(filmRow => filmRow.actors.map((_,1))).reduceByKey(_+_)
    val topActors = actorMovieCount.sortBy(_._2, ascending = false).map(_._1).take(5)

    val filmsWithTopActors = filmRows.withResult(topActors).filter{
      filmRowActors => {
        val (filmRow, actors) = filmRowActors
        filmRow.actors.toSet.intersect(actors.toSet).nonEmpty
      }
    }.map(_._1)
  }


  val columns = Seq(
    "title",
    "release",
    "locations",
    "funFacts",
    "productionCompany",
    "distributor",
    "director",
    "writer",
    "actor1",
    "actor2",
    "actor3"
  )

  case class InputFilmRow(title: String,
                          release: String,
                          locations: String,
                          funFacts: String,
                          productionCompany: String,
                          distributor: String,
                          director: String,
                          writer: String,
                          actor1: String,
                          actor2: String,
                          actor3: String)

  case class FilmRow( title: String,
                      release: String,
                      locations: Seq[String],
                      funFacts: String,
                      productionCompany: String,
                      distributor: String,
                      director: String,
                      writer: String,
                      actors: Seq[String])


  val parseFilmRows = (tuple: (String, Seq[InputFilmRow])) => {
    val (title, rows) = tuple
    val firstRow = rows.head
    val locations = rows.map(_.locations).distinct
    val actors = rows.flatMap(row => Seq(row.actor1, row.actor2, row.actor3)).distinct.filter(_ != "")
    FilmRow(
      firstRow.title,
      firstRow.release,
      locations,
      firstRow.funFacts,
      firstRow.productionCompany,
      firstRow.distributor,
      firstRow.director,
      firstRow.writer,
      actors)
  }

  def testFile(fileName: String): String = {
    Thread.currentThread().getContextClassLoader.getResource(fileName).toString
  }

}

Upcoming

• graphx support
• DAG viewer frontend attached to running process
• component / pipeline abstractions
• debug run mode with auto Try wrapped functions and trapped failures