Still working to reproduce this condition, if others are seeing something similar please describe it here so that we can narrow in on and hopefully reliably reproduce it...
In a m/r job, the majority of reducers are fairly quickly successful and complete. Then a small subset slow to an absolute crawl. There are some overflow rows - but not an abnormal amount as compared to other successful reducers. It indexes locally, and I notice in the main index directory there are around 14k files which I imagine is an important clue.
I changed two things at once to get the job to complete successfully in a reasonable amount of time. I limited the row size to max of 250k records and added a multiplier of 4.