Hadoop streaming creates 2 mappers when the input has only one block

If a streaming job is run against input that is less than 2 blocks, 2 mappers will be created, both operating on the same split, both producing (duplicate) output. In some cases the second mapper will consistently fail. I've not seen the failure with input less than 10 bytes or more than a couple MB. I have seen it with a 4kB input.