You can have a separate connection pool for 'cache' requests. You shouldn't have too many PG connections open anyway, on the order of O(num of CPUs).