A Deep Dive into KEDA
If you’ve spent any time in the Kubernetes ecosystem, you’re likely familiar with the Horizontal Pod Autoscaler (HPA). It’s the reliable workhorse of scaling, but it has a bit of a "one-track mind." By default, HPA makes decisions based on resource metrics like CPU and Memory.
But what happens when your traffic isn't CPU-bound? What if your app is sitting idle until a message hits a RabbitMQ queue, or a new blob is uploaded to Azure Storage?
That’s where KEDA (Kubernetes Event-Driven Autoscaling) steps in. Think of it as the "brain" that tells Kubernetes how to react to the outside world.
What exactly is KEDA?
KEDA is a single-purpose, lightweight component that can be added to any Kubernetes cluster. It acts as an operator that bridges the gap between external event sources and the native Kubernetes HPA.
While Kubernetes is great at managing containers, it doesn’t natively "speak" the language of Kafka, Prometheus, or AWS SQS. KEDA provides those translators (called Scalers) to trigger scaling based on the actual demand of your event-driven applications.
Why Use KEDA? (The "Killer Features")
Scale to Zero: This is the holy grail of cost-saving. Standard HPA can only scale down to 1 replica.
KEDA can scale your deployment down to 0 when there’s no work to do, and "activate" it back to 1 the moment an event arrives.
Massive Library of Scalers: Whether you're using Redis, PostgreSQL, GCP Pub/Sub, or even a basic Cron schedule, KEDA likely has a built-in scaler for it.
HPA Compatibility: KEDA doesn't replace the HPA; it feeds it. It acts as a metrics server, so you get the benefits of KEDA’s event-awareness with the battle-tested stability of the native Kubernetes autoscaler.
Vendor Agnostic: You can run KEDA on-prem, in AKS, EKS, or GKE. It gives you a "Serverless" experience on any cloud without the vendor lock-in of functions-as-a-service (FaaS).
How It Works: The Three Pillars
KEDA functions using three primary components:
Scalers: These are the integrations with external systems (e.g., the Kafka Scaler). They monitor the event source and report the "lag" or "load."
Metrics Adapter: This takes the data from the Scaler and converts it into a format the Kubernetes HPA can understand.
Controller: This is the "manager" that creates and updates the HPA objects and handles the scale-to-zero/scale-from-zero logic.