A Kubernetes app pulls Pub/Sub messages and writes to Filestore; a single pod cannot process messages in real time and messages back up. The processing is I/O-intensive. How should you scale processing?

The correct solution is to configure Kubernetes autoscaling based on the subscription/numundeliveredmessages Pub/Sub metric. This directly addresses the problem of messages backing up by scaling out the processing pods when the number of undelivered messages increases. CPU-based autoscaling (--cpu-percent=50) is less effective here because the processing is I/O-intensive, meaning CPU utilization might not be a direct indicator of backlog. Cluster autoscaling (--enable-autoscaling) scales the underlying nodes, not the application pods themselves, and wouldn't directly respond to the Pub/Sub backlog. While subscription/pushrequestlatencies could indicate a problem, numundeliveredmessages is a more direct and actionable metric for scaling out consumers to clear the backlog.