kedacore/keda

The problem

Kubernetes' Horizontal Pod Autoscaler scales on CPU and memory, which is a poor proxy for work that arrives as messages: a consumer can be idle on CPU while a queue backs up, or burn CPU while the queue is empty. Teams end up writing custom controllers or cron-based workarounds. KEDA adds an event-driven scaling layer that keeps the standard HPA and supplies it metrics scraped from external event sources, including the ability to scale a workload to zero when there is nothing to process.

What's different

Compared with the built-in HPA alone, the difference is the metric source: instead of resource utilisation, KEDA reads queue depth, lag, or similar signals through scalers and exposes them to the HPA. Compared with Knative-style serverless platforms, KEDA does not replace your workload runtime; it wraps ordinary Deployments and Jobs, which makes it far less invasive. The zero-to-zero capability and a ScaledJob path for batch work are the parts that actually change architecture. The README is light on internals, so details belong to the docs site rather than this file.

Who it's for

Platform and SRE teams running queue consumers, stream processors, or batch jobs on Kubernetes, especially where idle services should shrink to nothing and burst traffic should scale out on backlog instead of CPU. It suits both cloud and edge clusters. Teams that only need CPU-based scaling should stay with plain HPA and skip another controller. Teams already invested in Knative or a managed serverless runtime may find the overlap not worth it.

Cost of adoption

You install a cluster-wide operator, which needs CRDs, RBAC, and a metrics server role, so changes go through cluster admins. Each event source needs its own scaler configuration and credentials. KEDA is Apache-2.0, a CNCF graduated project, created in 2019, so it is mature rather than experimental. The main risks are operational: tuning polling intervals and scale thresholds correctly, and understanding how zero-replica behaviour interacts with your ingress and readiness.

Verdict

This is an established CNCF graduated project, not a newcomer: 10.6k stars since 2019 and steady monthly attention (+124 stars in a recent month) point to sustained use rather than a spike. The Hacker News footprint is negligible, which fits infrastructure tooling that spreads through operations teams rather than front-page buzz. The README is mostly links, so evaluate from keda.sh before committing. If your workloads are event-driven, it removes real custom plumbing; if they are plain request-serving services, the HPA is already enough.