Queues are excellent shock absorbers when producers and consumers truly operate on different clocks. They become a trap when every spike is merely deferred into a deeper backlog that still must drain under the same dependency limits.
Before you introduce another broker, measure whether the consumer can ever catch up at peak arrival rate. If not, the queue only buys time to page people — it does not buy capacity. Prefer back-pressure, shedding, or earlier admission control when the consumer ceiling is known.
When buffering is warranted, publish the drain SLO alongside the enqueue path. Teams that cannot state how long a message may wait are not running a queue strategy; they are running a hope strategy.
