So for now, I solved this problem of doing a map-side join by partitioning my sequence files (actually, map files) such that they are range partitioned on the key. For every new Spark partition (as returned by mapPartition), I look at the very first key and based on that key, I open a MapFile.Reader on the right map file. The successive keys are guaranteed to be in that same map file, and hence can be served by the already opened reader.
While this works, I would like to explore overriding HadoopRDD as per Reynold's suggestion, as it would make for a cleaner implementation. Thanks for your suggestions. Ameet On Mon, Oct 21, 2013 at 2:12 PM, Reynold Xin <[email protected]> wrote: > Maybe you can override HadoopRDD's compute method to do that? > > > On Mon, Oct 21, 2013 at 8:16 AM, Ameet Kini <[email protected]> wrote: > >> Right, except both my sequence files are large and so doing a "collect()" >> and then broadcasting one of them would be costly. Since I have two large >> sorted sequence files with a one-to-one relationship among the keys, I need >> to perform the "merge" portion of a good old "sort-merge" join. And it is >> actually a very simple merge, since each key is unique within the file. >> >> I was looking at the mapPartitions API: >> def mapPartitions[U](f: (Iterator[T]) => Iterator[U], preservesPartitioning: >> Boolean)(implicit arg0: ClassManifest[U]): >> RDD<http://spark.incubator.apache.org/docs/latest/api/core/org/apache/spark/rdd/RDD.html> >> [U] >> >> If somehow the function f has access to the underlying partition >> information (e.g., HadoopPartition.inputSplit), then it could open a reader >> on the actual hdfs file corresponding to that inputSplit, and manually do >> the join. But looks like HadoopPartition is declared private. Is there some >> other way to figure out which underlying HDFS file corresponds to the >> partition being iterated upon in mapPartitions? >> >> Ameet >> >> >> >> >> On Mon, Oct 21, 2013 at 12:54 AM, Reynold Xin <[email protected]> wrote: >> >>> How about the following: >>> >>> val smallFile = sc.sequenceFile(....).collect() >>> val largeFile = sc.sequenceFile(...) >>> >>> val small = sc.broadcast(smallFile) >>> largeFile.mapPartitions { iter => >>> // build up a hash table for small. called it smallTable >>> iter.filter(row => smallTable.contains(row.joinKey)).map { row => >>> join smallTable.get(row.joinKey) with row itself >>> } >>> } >>> >>> >>> >>> >>> On Fri, Oct 18, 2013 at 2:22 PM, Ameet Kini <[email protected]> wrote: >>> >>>> Forgot to add an important point. My sequence files are sorted (they're >>>> actually Hadoop map files). Since they're sorted, it makes sense to do a >>>> fetch at the partition-level of the inner sequence file. >>>> >>>> Thanks, >>>> Ameet >>>> >>>> >>>> On Fri, Oct 18, 2013 at 5:20 PM, Ameet Kini <[email protected]>wrote: >>>> >>>>> >>>>> I've seen discussions where the suggestion is to do a map-side join, >>>>> but haven't seen an example yet, and can certainly use one. I have two >>>>> sequence files where the key is unique within each file, so the join is a >>>>> one-to-one join, and can hence benefit from a map-side join. However both >>>>> sequence files can be large, so reading one of them completely in the >>>>> driver and broadcasting it out would be expensive. >>>>> >>>>> I don't think there is a map-side join implementation in Spark but >>>>> earlier suggestions have been to write one using mapPartitions on one of >>>>> the operands as the outer loop. If that is the case, how would I fetch the >>>>> split corresponding to the keys in the outer's partition. I'd prefer to do >>>>> a fetch-per-partition rather than a fetch-per-tuple. >>>>> >>>>> In any case, some feedback, and preferably, an example of a map-side >>>>> join without broadcasting would help. >>>>> >>>>> Thanks, >>>>> Ameet >>>>> >>>> >>>> >>> >> >
