Throughput measures how many requests an application receives or serves each second. In HTTP deployments, requests per second (RPS) provides a practical view of demand and replica capacity. section 6.2.1
Use RPS to plan capacity: measure how many requests one replica can handle, divide current demand by that capacity, and provision the resulting number of replicas. The same metric can drive Kubernetes horizontal pod autoscaling. section 6.2.1
RPS is especially useful when:
- CPU utilization does not represent demand because an HTTP service spends substantial time waiting on databases or other external services. section 6.2
- A tested replica capacity provides confidence that enough replicas can be provisioned under load. section 6.2.1
- Traffic enters through an Ingress, which sees HTTP requests and can expose the external metric. section 6.2.1
Scaling an HTTP deployment
Expose a Deployment through a NodePort Service and Ingress, then configure an HPA with an external request-count metric. Generate traffic with ApacheBench and observe the HPA increase replicas when demand exceeds the target.
