Is the pragmatic solution to return 503 and have clients back off. Or, if possible make latency a feature (embrace the queue!). For service to service internal

Is the pragmatic solution to return 503 and have clients back off. Or, if possible make latency a feature (embrace the queue!). For service to service internal stuff e.g. something like a request to hard delete something, this can always be a queue. And obviously you can scale up as the queue backs up. I do love the maths tho!