@takeshitakenji It isn't so much the number per se.  A large instance can be expected to have a large queue.  The problem is that failed entries are staying in the queue, and every time the queue executes, be it on OpprotunisticQM or the daemons (STOMP is different though), all the failed entries are tried before the new ones.

This is why Wrongthink ended up so far behind us by the way - they had days turned to weeks turned to a month of dead queue entries, that were being failed and executed before a new one, which would then affect the HMAC and time it's visible, and it snowballs from there.