Add research note: KEDA — Kubernetes Event-Driven Autoscaling

research/keda.md

@@ -0,0 +1,68 @@
+---
+title: "KEDA — Kubernetes Event-Driven Autoscaling"
+author: Ruben Koster (@rkoster)
+date: 2026-07-02
+tags: [autoscaling, runtime-lifecycle]
+cf_areas: [diego]
+status: draft
+sources:
+  - https://keda.sh/
+  - https://github.com/kedacore/keda
+---
+
+## Summary
+
+KEDA (Kubernetes Event-Driven Autoscaling) is a [CNCF graduated project](https://www.cncf.io/projects/keda/)
+that adds event-driven and queue-depth-based autoscaling to Kubernetes, including true
+scale-to-zero and scale-from-zero. It sits alongside the [Kubernetes HPA](https://kubernetes.io/docs/concepts/workloads/autoscaling/horizontal-pod-autoscale/),
+extending it with external metric sources rather than replacing it. KEDA is the de facto
+standard for event-driven scaling on Kubernetes and ships [70+ scalers](https://keda.sh/docs/latest/scalers/)
+([source](https://github.com/kedacore/keda/tree/main/pkg/scalers)) for common messaging
+and queuing systems.
+
+## Key findings
+
+- **Three-component architecture**: KEDA Operator (watches ScaledObject CRDs and drives
+  scaling decisions), Metrics Server (feeds external metrics to the Kubernetes HPA), and
+  Scalers (connectors to event sources that pull metric data). Admission webhooks validate
+  ScaledObject configurations at admission time.
+- **Two CRDs for two workload shapes**: `ScaledObject` targets long-running
+  Deployments/StatefulSets and manages replica count continuously. `ScaledJob` targets
+  batch workloads — it creates Kubernetes Jobs on demand and cleans them up on completion,
+  avoiding the problem of long-running workers accumulating in a `Terminating` state.
+- **Activation vs. scaling thresholds**: KEDA separates the threshold for waking a
+  scaled-to-zero workload (`activationThreshold`) from the threshold that drives ongoing
+  scaling (`threshold`). This prevents false-wake oscillation on low-volume queues.
+- **Scaler catalog**: 70+ built-in scalers covering RabbitMQ, Kafka, Redis (Lists and
+  Streams), NATS JetStream, Temporal, Prometheus custom metrics, and cron schedules.
+  TriggerAuthentication CRDs handle credential management for event source access,
+  supporting cloud IAM, Kubernetes secrets, and Vault.
+- **Pausing and forced activation**: Individual ScaledObjects can be paused (freeze
+  at current replicas, scale to zero, or block scale-in/out independently) and
+  force-activated (bypass the activation threshold to immediately scale from zero). This
+  gives operators fine-grained control beyond simple min/max bounds.
+- **Long-running execution handling**: KEDA provides two patterns for workloads that
+  must not be interrupted mid-task: container lifecycle hooks that delay termination on
+  SIGTERM until the current batch completes, and ScaledJob which models each task as a
+  Job that runs to completion and exits cleanly.
+
+## CF relevance
+
+KEDA shows a mature, production-proven approach to the autoscaling problem that
+HTTP-biased platforms have not needed to solve. The key gap it addresses — scale
+workloads to zero when a queue is empty and wake them when messages arrive — is directly
+relevant to agent workloads where tasks are discrete, arrival is unpredictable, and idle
+compute is waste. The ScaledJob pattern is particularly relevant for agent sandboxes: each
+tool invocation or sub-task could be modelled as a job that runs to completion rather than
+a long-lived process waiting for work.
+
+## Open questions
+
+- KEDA operates at cluster scope; a multi-tenant platform would need space-scoped scaling
+  boundaries — how would that work without a per-space KEDA operator?
+- TriggerAuthentication maps naturally to service bindings — but does it support
+  SPIFFE/SVID-based workload identity as a credential source, or only static secrets and
+  cloud IAM?
+- For agent workloads that hold expensive state in memory (loaded models, cached context),
+  scale-to-zero means losing that state — how do KEDA users handle warm pools or
+  pre-warming to amortize cold-start cost?