Calling Scalding from inside your application

Starting in scalding 0.12, there is a clear API for doing this. See Execution[T], which describes a set of map/reduce operations that when executed return a Future[T]. See the scaladocs for Execution. Below is an example.

val job: Execution[Unit] =
  TypedPipe.from(TextLine("input"))
    .flatMap(_.split("\\s+"))
    .map { word => (word, 1L) }
    .sumByKey
    .writeExecution(TypedTsv("output"))
// Now we run it in Local mode
val u: Unit = job.waitFor(Config.default, Local(true))

// Or for Hadoop:
val jobConf = new JobConf
val u: Unit = job.waitFor(Config.hadoopWithDefaults(jobConf), Hdfs(jobConf, true))
// If you want to be asynchronous, use run instead of waitFor and get a Future in return

For testing or cases where you aggregate data down to a manageable level, .toIterableExecution on TypedPipe is very useful:

val job: Execution[Iterable[(String, Long)]] =
  TypedPipe.from(TextLine("input"))
    .flatMap(_.split("\\s+"))
    .map { word => (word, 1L) }
    .sumByKey
    .toIterableExecution
// Now we run it in Local mode
val counts: Map[String, Long] = job.waitFor(Config.default, Local(true)).toMap

To run an Execution as a stand-alone job, see:

1. ExecutionApp Make an object MyExJob extends ExecutionApp for a job you can run like a normal java application (by using java on the classname).
2. ExecutionJob - use this only if you have an existing tooling around launching scalding.Job subclasses.

Some rules

1. When using Execution NEVER use .write or .toPipe (or call any method that takes an implicit flowDef). Instead use .writeExecution, .toIterableExecution, or .forceToDiskExecution. (see scaladocs).

2. Avoid calling .waitFor or .run AS LONG AS POSSIBLE. Try to compose your entire job into on large Execution using .zip or .flatMap to combine Executions. waitFor is the same as run except it waits on the future. There will be 1 (and should be only 1) calling to .waitFor or .run in each Execution App/Job.

3. Only mutate vars or perform side effects using .onComplete. If you run the result of onComplete, your function you pass will be run when the result up to that point is available and you will get the Try[T] for the result. Avoid this if possible. It is here to deal with external IO, or existing APIs, and designed for experts that are comfortable using .onComplete on scala Futures (which is all this method is doing under the covers).

Running Existing Jobs Inside A Library

We recommend the above approach to build composable jobs with Executions. But if you have an existing Job, you can also run that:

Working example:

WordCountJob.scala

class WordCountJob(args: Args) extends Job(args) {
  TextLine(args("input"))
    .read
    .flatMap('line -> 'word) { line: String => line.split("\\s+") }
    .groupBy('word) { _.size }
    .write(Tsv(args("output")))
}

Runner.scala

object Runner extends App {
  val hadoopConfiguration: Configuration = new Configuration
  hadoopConfiguration.set("mapred.job.tracker","hadoop-master:8021")
  hadoopConfiguration.set("fs.defaultFS","hdfs://hadoop-master:8020")

  val hdfsMode = Hdfs(strict = true, hadoopConfiguration)
  val arguments = Mode.putMode(hdfsMode, Args("--input in.txt --output counts.tsv"))

  // Now create the job after the mode is set up properly.
  val job: WordCountJob = new WordCountJob(arguments)
  val flow = job.buildFlow
  flow.complete()
}

And then you can run your App on any server, that have access to Hadoop cluster