After debugging this, I found that the map-reduce output collection is written to the master, but if you connected to the replica set with SLAVE_OK=true, you may very well be reading from a slave node before the output collection has been synced there, which will indeed give you empty results.

The workaround I put in my code is to ensure that the output collection never has the SLAVE_OK=true option, no matter what connection options were set. However, I suggest putting this in the MapReduceOutput code. This could be added to the constructor:

_coll = db.getCollection( _collname );
if (_coll != null) _coll.setOptions(_coll.getOptions() & ~Bytes.QUERYOPTION_SLAVEOK);

I have tested similar code with my applications, and it has indeed made this bug go away!