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.