Queueing connections and avoiding the "Exceeded concurrency limit" error

[mtvx]

With backlog of 2048 and limit-concurrency of 20 we're seeing Exceeded concurrency limit errors pop up way before the backlog should be filled.

My assumption would be that if all tasks are running, the connections get queued (up to backlog limit) and then get routed to the workers.

However, that does not seem to be the case (though I don't have metrics on this, at this point it's just an assumption). Looking at the code, the if-block that generates the Exceeded concurrency limit does not seem to consider backlog at all?

Am I misunderstanding something, how can I get connections first queued, then routed to available workers/tasks as they free up?