KEDA: Scaling Kubernetes Workloads on Real Demand, Not Just CPU and Memory

In this session, Vaibhav Vardhiya, Senior DevOps Engineer at Ksolves India Limited, addresses a limitation that becomes increasingly visible as Kubernetes workloads grow in variety and complexity: CPU and memory metrics are useful signals, but they are not always the right signals for autoscaling decisions.

For many real-world workloads, CPU utilization and memory consumption do not reliably reflect actual demand. A queue-driven microservice processing messages from Kafka or RabbitMQ may sit at low CPU utilization while thousands of messages accumulate in the backlog. A batch job may need to spin up instantly when a scheduled trigger fires, regardless of what the current resource metrics show. A machine learning inference service may need to scale to zero overnight and come back up immediately when requests arrive in the morning. In all of these cases, scaling on CPU and memory alone leaves infrastructure under-responsive, inefficient, or unnecessarily expensive.

KEDA, Kubernetes Event-Driven Autoscaling, was built to solve this. It works alongside the native Kubernetes Horizontal Pod Autoscaler to enable scaling decisions based on the signals that actually drive the workload: queue depth, stream lag, scheduled events, and external triggers from systems like Kafka, RabbitMQ, Azure Service Bus, and more. This webinar covers how KEDA works in practice, walking through ScaledObjects and ScaledJobs, scale-to-zero and scale-from-zero behavior, event-based triggers for both applications and batch workloads, and practical applications across queue-driven microservices, streaming pipelines, scheduled jobs, and ML inference. The session also covers how KEDA improves infrastructure responsiveness while reducing unnecessary cloud costs by ensuring workloads only consume resources when demand actually exists.

If you are running Kubernetes workloads and looking to build more responsive, elastic, and cost-efficient infrastructure, this session provides practical insights into scaling based on what is actually happening in your systems.

Key Takeaways

• Why CPU and memory metrics are not always sufficient signals for effective Kubernetes autoscaling.
  
• How KEDA enables event-driven scaling using real demand signals such as queue depth and stream lag.
  
• Understanding KEDA, HPA, ScaledObjects, and ScaledJobs and how they work together.
  
• How to scale workloads to zero and back up automatically based on actual demand.
  
• Using event-based triggers for queue-driven microservices, streaming pipelines, scheduled jobs, and ML inference.
  
• How KEDA improves infrastructure responsiveness while reducing unnecessary cloud costs.