KEDA — Event-Driven Autoscaling on Kubernetes

KEDA extends the Horizontal Pod Autoscaler with 60+ built-in scalers for external event sources — CloudWatch, Kafka, SQS, Prometheus, and more. It does not replace HPA; it feeds HPA with metrics from sources that the standard Kubernetes metrics pipeline cannot reach.

Key Takeaways

• KEDA consists of three components: operator (polls external sources), metrics server (exposes external.metrics.k8s.io), and admission webhooks (validates ScaledObject configs)
• ScaledObject maps a Deployment to an event source; ScaledJob creates Jobs per event — both are CRDs
• On EKS, KEDA uses IRSA for AWS scaler authentication — no long-lived credentials needed
• Scale-up is fast (one polling interval, default 30s); scale-down is slow (cooldown period, default 300s) — this prevents flapping
• Scale-to-zero is supported but requires explicit opt-in and is best suited for batch/job workloads

Quick Overview

The keda-operator watches ScaledObject resources, creates backing HPAs, and polls external sources at a configurable interval. Metric values are served through keda-operator-metrics-apiserver via the Kubernetes external metrics API — the HPA consumes them transparently. Admission webhooks validate configurations and prevent duplicate ScaledObjects from targeting the same Deployment.

A typical AWS CloudWatch scaler queries ALB RequestCountPerTarget and scales a Deployment when the value exceeds a threshold. During load testing with hey, scale-up from 1 to 10 pods completes within ~3 minutes, and scale-down takes ~5 minutes of cooldown.

---

Read the full guide: KEDA — Kubernetes Event-Driven Autoscaling → — includes complete YAML examples, CloudWatch Metrics Insights expressions, IRSA setup with Terraform, scaler reference table, production monitoring with Prometheus, and a failure mode runbook.