Kubernetes Horizontal Pod Autoscaler (HPA) and KEDA: Principles, Limitations, and Implementation

This article provides a comprehensive analysis of Kubernetes horizontal pod autoscaling (HPA) and KEDA (Kubernetes Event-driven Autoscaling). It begins by explaining HPA's fundamental principles, including its evolution through three major versions (autoscaling/v1, v2beta1, and v2beta2) and support for four metric types: Resource, Object, Pods, and External. The article details HPA's working algorithm, which calculates desired replicas using the formula: desiredReplicas = ceil[currentReplicas * (currentMetricValue / desiredMetricValue)]. It emphasizes that HPA requires defined resource requests and metrics-server installation to function properly.

The article then discusses HPA's limitations, including its inability to scale to zero, potential performance bottlenecks with single-threaded architecture, and issues with multi-container pod resource utilization. It provides a practical example of deploying an nginx service with HPA configured to scale based on CPU utilization thresholds.

The second major section introduces KEDA as an enhancement to HPA, explaining that KEDA is an event-driven autoscaler that supports scaling to zero and provides 30+ built-in scalers for various external sources. The article explains KEDA's core components: Metrics Adapter, HPA Controller, and Scaler, and how KEDA extends HPA's capabilities by handling the scale-to-zero functionality.

Implementation details include KEDA deployment requirements (Kubernetes 1.16+), configuration through ScaledObject resources, and various trigger types including Prometheus metrics, CPU utilization, and cron-based scheduling. The article provides code examples for both HPA and KEDA configurations.

Finally, the article discusses practical use cases for HPA and KEDA, including big data batch processing, event-driven architectures, conversion services, and cost optimization through traffic-based scaling. It concludes with references to official documentation and community resources.