This code allows parallel loading of data from S3 to Spark RDD. Support multiple paths to load from. Based on http://tech.kinja.com/how-not-to-pull-from-s3-using-apache-spark-1704509219

spark-create-rdd-from-s3-parallel.scala

val s3Paths = "s3://yourbucket/path/to/file1.txt,s3://yourbucket/path/to/directory"
val pageLength = 100
val key = "YOURKEY"
val secret = "YOUR_SECRET"

import com.amazonaws.services.s3._, model._
import com.amazonaws.auth.BasicAWSCredentials
import com.amazonaws.services.s3.model.ObjectListing
import scala.collection.JavaConverters._
import scala.io.Source
import java.io.InputStream
import org.apache.spark.rdd.RDD

def s3 = new AmazonS3Client(new BasicAWSCredentials(key, secret))
var inputLinesRDD_raw:RDD[String] = sc.emptyRDD[String]
s3Paths.split(",").foreach{ s3Path =>
  val regex = """(?i)^s3://([^/]+)/(.*)""".r
  val bucket = regex.findFirstMatchIn(s3Path).map(_ group 1).getOrElse(null)
  val prefix = regex.findFirstMatchIn(s3Path).map(_ group 2).getOrElse(null)
  println("Processing s3 resource: bucket '%s', prefix '%s'".format(bucket, prefix))
  @transient val request = new ListObjectsRequest()
  request.setBucketName(bucket)
  request.setPrefix(prefix)
  request.setMaxKeys(pageLength)
  @transient var listing = s3.listObjects(request)
  var proceed = true
  while (proceed){
    if (listing.getObjectSummaries.isEmpty){
      proceed = false
    }else{
      @transient val s3FileKeys = listing.getObjectSummaries.asScala.map(_.getKey).toList
      val inputLines = sc.parallelize(s3FileKeys).flatMap { key => Source.fromInputStream(s3.getObject(bucket, key).getObjectContent: InputStream).getLines }
      inputLinesRDD_raw = inputLinesRDD_raw.union(inputLines)      
      listing = s3.listNextBatchOfObjects(listing)
    }
  }
}

// TODO do something with inputLinesRDD_raw

@snowindy @hamishdickson If my Input data is ORC and if am just interested in fetching out say 20 columns out of 120 columns present in the files (columnar data access) , how does this work ?

Would love to see the python port

@kakkarut @narayana1043

Below is the code I used to UNLOAD data from Redshift to S3, then read S3 into Spark. You could refer to Line 62~72,

https://github.com/skinheadbob/trinity/blob/master/aws/rsutil.py

The idea is the same: parallelize( s3_keys ).flatMap( lambda: read_lines(s3_key) ). Hope it helps.

FYI @wojtekk and @dineshjlabs and anyone else that has this problem - I managed to fix this by wrapping AmazonS3Client in a new class like so (I'm using groovy for this project)

import com.amazonaws.auth.BasicAWSCredentials
import com.amazonaws.regions.Regions
import com.amazonaws.services.s3.AmazonS3Client

class SerializableAmazonS3Client extends AmazonS3Client implements Serializable {
    public SerializableAmazonS3Client(BasicAWSCredentials credentials) {
        super(credentials)

       // I'm also setting my region here because it's convenient
        super.setRegion(Regions.EU_WEST_1)
    }
}

I haven't got time to test it, but I guess the scala would just be

class SerializableAmazonS3Client(creds: BasicAWSCredentials) extends AmazonS3Client(creds) with Serializable

This worked for me. Thanks @hamishdickson !