Set varys.master.numInstances > 1 and run the BroadcastService example with a large file (e.g., 32MB) should reproduce the bug. I suspect it to be a race condition somewhere.
varys.master.numInstances determines the number of MasterActor, which should've helped with processing too many messages. Setting it to 1 isn't really that bad because most of the time is spent in (de)serialization and not in actual action.