Scaling
Choose what to scale and the signal that should drive it. Keep one controller responsible for each replica or resource field.
Horizontal scaling
Section titled “Horizontal scaling”KEDA scalers read queues and other event sources. KPA provides a horizontal control loop; Scaling Groups share a replica budget.
Choose an event-source scaler. Replica and job scaling explains the workload choice.
HTTP scaler scales from request rate or concurrency and supports activation from zero. Envoy HTTP scaler uses an existing Envoy metrics pipeline and requires warm replicas.
Choose a routing integration, including Gateway API and Istio, or use the inference routing.
OpenTelemetry
Section titled “OpenTelemetry”OpenTelemetry scaler scales replicas or jobs from application metrics. Start with an OTel metric, then choose an ingestion method. Recipes cover Prometheus migration, Ingress NGINX and vLLM.
Predictive and scheduled scaling
Section titled “Predictive and scheduled scaling”Predictive scaler uses historical demand. ScalingPolicy changes scaling configuration during scheduled windows. Compare forecasts, Cron and policies.
Vertical scaling
Section titled “Vertical scaling”Pod Resource Autoscaler changes resources from utilization. Pod Resource Profiles applies explicit resource profiles. Compare their requirements and ownership.
Multi-cluster
Section titled “Multi-cluster”Register member clusters, then distribute replicas or jobs. vCluster scaling centralizes scaling for virtual-cluster workloads.
Custom targets and node capacity
Section titled “Custom targets and node capacity”ScaleAdapter exposes custom replica fields. Node capacity buffers prepare capacity using Karpenter, Cluster Autoscaler or GKE.
Related tasks
Section titled “Related tasks”Insights recommends resource changes to review. Operate covers fleet and tenant management. Use Best Practices to tune and validate production behavior, or Reference for exact configuration fields.