In the select_fun for the intersection_op (i.e. AND) it would stop
once one of it's iterators had been exhausted but would leave the
other one as is, messages and all. So in the case where one side is
empty and other side has a bunch of results (or this query is run a
lot, memory usage = result size * # of queries) the results just sit
on a process that waits indefinitely for something to pull the results
off it's mailbox. The reason this doesn't occur if you use a PB conn
for each transaction is because he PB conn is the parent of these
processes and they are all linked. If it dies they all die.
The solution is to make sure all iterators are exhausted.