I have always used the load average to decide whether to scale a cluster of Docker Swarm or Kubernetes. Never had issues.

However, I have noticed Kubernetes' Cluster Autoscaler or AWS' AutoScaling Groups use CPU as the main metric to decide when to scale.

What's the reason for that? The load is a metric which serves as an overview of the overall status of the system, whereas CPU is just one of the metrics that impact the system.

Which metric are you using and why?