大清单FlatMap Java Spark

Question

I have a large list in JavaPairRDD<Integer, List<String>> and I want to do a flatMap to get all possible combinations of list entries so that I end up with JavaPairRDD<Integer, Tuple2<String,String>> . Basically if i have something like

(1, ["A", "B", "C"])

I want to get:

(1, <"A","B">) (1, <"A", "C">) (1, <"B", "C")

The problem is with large lists as what I have done is created a large list of Tuple2 objects by having a nested loop over the input list. Sometimes this list does not fit in memory. I found this, but not sure how to implement it in Java: Spark FlatMap function for huge lists

Answer 1

You may want to flatMap the list and then join the RDD on itself before filtering equal values:

JavaPairRDD<Integer, List<String>> original = // ...
JavaPairRDD<Integer, String> flattened = original.flatMapValues(identity());
JavaPairRDD<Integer, Tuple2<String, String>> joined = flattened.join(flattened);
JavaPairRDD<Integer, Tuple2<String, String>> filtered = 
    joined.filter(new Function<Tuple2<Integer, Tuple2<String, String>>, Boolean> () {
        @Override
        public Boolean call(Tuple2<Integer, Tuple2<String, String>> kv) throws Exception {
            return kv._2()._1().equals(kv._2()._2());
        }
    });

Answer 2

depends on how big of your datasets, in my job it usually have to process 100-200GB datasets used the FlatMap and flatMapToPair both is works properly for high intensive computation. example below

JavaPairRDD<Integer, List<String>>= DatasetsRDD.
    .flatMapToPair(x->{
    return xx;
    });

Also if your datasets are huge you could try to use spark persistance to disk

Storage Level   

    MEMORY_ONLY
    MEMORY_ONLY_SER 
    MEMORY_AND_DISK_SER 
    DISK_ONLY
    MEMORY_ONLY_2

References: https://spark.apache.org/docs/latest/rdd-programming-guide.html