You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@spark.apache.org by Stanley Burnitt <St...@huawei.com> on 2013/10/12 01:51:34 UTC
ReduceByKey OOME workaround: 'sample and subtract', but
rdd.subtract does not subtract.
Sampled RDDs contain the number of expected elements, but RDD.subtract(sample) does nothing; the returned RDDs have the same # of elements as the originalRDD.
Here is the code...
long count0 = resultRDD.count();
JavaPairRDD<KeyBytesWritable, AggregationWritable> r1 = resultRDD.sample(false, 0.50, 7);
JavaPairRDD<KeyBytesWritable, AggregationWritable> r2 = resultRDD.subtract(r1);
long count1 = r1.count();
long count2 = r2.count(); /* count0 == count2 ?! */
System.out.printf("RESULT-RDD.COUNT=%d ... COUNT1=%d ... COUNT2=%d \n", count0, count1, count2);
The printf output shows nothing was subtracted from the original rdd. Am I using the subtract method incorrectly?
The goal is to split a very large Pair-RDD containing 1 partition into a Pair-RDD containing many smaller partitions in hopes that this will fix OOM errors during a reduceByKey task.
After much tweaking of jvm GC parameters there is improvement, but eventually, the job just stalls out - no OOM, but no progress (jstack and jmap shows workers' new-gen is maxed out during RDD de-serialization).
Does anyone have any suggestions?
Thanks,
Stan