I think those two nodes have exhausted threadpool and have gone into a deadlock, given the number of single writes. I vaguely remember a similar issue that we fixed in a recent rc build. Let me upgrade one of the stalled nodes to see if that helps