Running scalaz-stream Processor inside Spark example

spark_scalaz-stream

import org.apache.spark._
import scalaz.stream._

/**
 * Simple proof of concept - fill an RDD from files that have been
 * processed by a scalaz-stream Process (in parallel). 
 */
object SparkScalazStream {

  def main(args: Array[String]) {

    val conf = new SparkConf().setAppName("Spark scalaz-stream test")
    val spark = new SparkContext(conf)

    val files = spark.parallelize(args.toSeq, args.length)

    val contents = files.flatMap { case f =>
      // assuming f exists on every node. would really read from HDFS...

      val in = scalaz.stream.io.linesR(f)
      val p = in //actually, some really complicated stream
                 //processing of in that relies on order, etc
      p.runLog //TODO MUST AVOID THIS!!!!
       .run
    }

    val lines = contents.map(_ => 1).reduce(_ + _)
    println("lines = " + lines)
    spark.stop()
  }
}

/*
* TODO Solve this problem:
* turn p ( a Process[Task,String] ) into a TraversableOnce[String]
* and let spark drive the state machine, rather than the Task
*/

The above works, but is not efficient:

• runLog collects the result in memory.
• run runs the Task (which drives the Process... so is another thread)
  There should be a better way...

hm, I think you can achieve whatever conversionyou want with process. You can either use Process.toTask or Process.step to lazily convert process to whatever primitive you want. But I would not recommend it. I think we have to hook to spark differently. Instead of files.flatMap, don't we have different combinator? Additionally, you can easily parallelize runing processes by using njoin or merge.mergeN from processes.

Thanks @pchlupacek. Would you mind elaborating?
I thought about implementing an iterator that steps through the Process on each next() and returns the emitted value somehow... but unsure of the details... I think this is what you meant with Process.step?
Why would you not recommend this?

I need to parallelise beyond a single host due to data size, so running Processes within spark seems a natural solution (I have a library of these and would like to lift them into spark - and later, I would also like to process data in RDDs with scalaz Processes via mapPartition). I don't know much about njoin or mergeN beyond reading the API just now, but I think these would be limited to a single host.

Since Iterator is a TraversableOnce, attempted this:
https://gist.github.com/florianverhein/2ed965bde7324cb73325