Iterate through a Java RDD by row

As mattinbits said in the comments, you want a map instead of a foreach, since you want to return values. What a map does basically is to transform your data: for each row of your RDD you perform an operation and return one value for each row. What you need can be achieved like this:

import org.apache.spark.api.java.function.Function;

...

SparkConf conf = new SparkConf().setAppName("PCA Example");
SparkContext sc = new SparkContext(conf);

JavaRDD<String> data = sc.textFile("clean-sl-mix-with-labels.txt",0).toJavaRDD();
JavaRDD<double[]> whatYouWantRdd = data.map(new Function<String, double[]>() {
    @Override
    public double[] call(String row) throws Exception {
        return splitStringtoDoubles(row);
    }

    private double[] splitStringtoDoubles(String s) {
        String[] splitVals = s.split("\\t");
        Double[] vals = new Double[splitVals.length];
        for(int i=0; i < splitVals.length; i++) {
            vals[i] = Double.parseDouble(splitVals[i]);
        }
        return vals;
    }
});

List<double[]> whatYouWant = whatYouWantRdd.collect();

So that you know how Spark works, you perform actions or transformations on your RDD. For instance, here we are transforming our RDD using a map function. You need to create this function yourself, this time with an anonymous org.apache.spark.api.java.function.Function which forces you to override the method call, where you receive a row of your RDD and return a value.

Just because it's interesting to compare the verboseness of the Java vs Scala API for Spark, here's a Scala version:

import org.apache.spark.{SparkContext, SparkConf}

class example extends App {
  val conf = new SparkConf().setMaster("local").setAppName("Spark example")
  val sc = new SparkContext(conf)

  val inputData = List(
    "1.2\t2.7\t3.8",
    "4.3\t5.1\t6.3"
  )

  val inputRDD = sc.parallelize(inputData)
  val arrayOfDoubleRDD = inputRDD.map(_.split("\t").map(_.toDouble))
}