> So, should we favor the draining of defer_linger_chain as much workers > as necessary like the current patch, or should we have as few workers > as possible and not start new workers in loops with no effect on > defer_linger_chain?
I think the fewer workers option could lead to hard to debug (from an end user POV) intermittent problems with clients far back in the queue who see a FIN delayed by someone ahead in the queue. They may send subsequent requests in the meantime that will spin.