You are managing a fleet of Compute Engine instances in a Managed Instance Group (MIG). The application processes messages from a Cloud Pub/Sub topic. Traffic is highly variable. You want the MIG to scale out when there is a backlog of messages and scale in when the queue is empty. How should you configure the autoscaler?

For worker nodes processing a queue, CPU is often a poor scaling metric (e.g., if workers are blocked waiting for a database, CPU is low, but the queue is growing). The best practice is to autoscale based on the actual backlog. GCP allows MIGs to autoscale based on any Cloud Monitoring metric, including `pubsub.googleapis.com/subscription/num_undelivered_messages`.