The scenario is that one thread begins to index a document (in this case, passing in an indexable map). It calls startDocument. Another thread wakes up and tells the engine to do a dump. That checks the indexing queue for a new doc, determines there isn't one (the AbstractPipelineImpl is already holding on to it), and goes ahead and does a realDump. That sticks a new InvFileMemoryPartition into the pipeline. Now the first thread calls handleField, sends its data down to the InvFileMemoryPartition, and it throws an NPE when it tries to get a dockey (which is probably for the best since otherwise it'd be pretty screwed up!).