Watching The Wrong Meter

A service mostly waits on a slow third-party API. During an incident, latency climbs and requests pile up, but autoscaling never fires because CPU sits at 15%. Which scaling signal fixes this?

Keep CPU as the scaling metric but drop the target from its current value to 10 percent so the group scales out well before saturation

Scale on outbound network bytes, which rises in step with the request volume being served

Scale on in-flight requests or queue depth per instance, which tracks the resource that is actually exhausted

Scale on memory utilization, since every request waiting on the third party holds its response buffers and connection state until it completes