For an ECS service that processes messages from a queue, CPU usage doesn’t always tell us whether we’re keeping up. The tasks can be mostly idle while the backlog grows, or briefly use a lot of CPU while the queue is still draining normally.

What metric has worked best for scaling this kind of service: queue depth, age of the oldest message, CPU, or a combination? I’m also curious how people avoid scaling up and down too aggressively when traffic arrives in bursts.