Queues should not get filled very often, most probably if queue gets full then there's some other problems that should be solved.

Most common reason for queues getting filled will be requests timing out with long timeouts. This is something that can be somewhat handled with QoS.

Other reasons can include mods that do something really stupid like sending request from global dig handler etc. and these reasons are not something that should be fixed with QoS.

It would probably be better to have bit smaller queues, for example if parallel requests is set to 16 then with current defaults:

• priority 1 queue size is 256 requests + 64 requests in engine queue
• priority 2 queue size is 192
• priority 3 queue size is 128

Assuming worst case scenario with 640 requests queued or reading reply and probable cause which is timeouts.
Default request timeout is 3 seconds, if all requests are going to time out processing whole queue takes at least 156 seconds to complete (taking processing order and priority limits into account).
Such situation can be very likely if some network service that is called frequently goes down and stays down for some time.
Very related: #2