First KEDA Setup

Event-driven scaling.

Overview

Your first KEDA setup is the moment Kubernetes scaling moves from CPU/memory thresholds to event-driven signals. KEDA scales workloads based on external metrics (Kafka consumer lag, RabbitMQ queue depth, SQS message count, Prometheus queries, 50+ scalers), supports scale-to-zero for idle workloads, and integrates with the standard HPA mechanism so the K8s ops surface stays familiar. The discipline locks in patterns the team will reuse for every queue-driven and event-driven workload.

• Event-driven scaling. Scale based on Kafka consumer lag, RabbitMQ depth, SQS message count; matches scaling to actual demand rather than CPU as a proxy.
• Scale to zero. Idle workloads consume nothing; the cost savings recur on every idle minute.
• Many scalers. 50-plus event-source scalers ship with KEDA; covers most queue and event-stream sources.
• HPA integration plus Helm install. KEDA generates HPAs under the hood; standard chart-based install; the K8s surface stays familiar.

The approach

The practical approach is helm install kedacore/keda for the standard install, define ScaledObject CRDs that describe what scales and based on what signal, set cooldown periods to prevent thrashing on transient signal dips, set explicit min and max replicas so scale-to-zero or runaway scale do not surprise the team, and document the per-workload scaling trigger committed to the deployment repo.

• helm install kedacore/keda. Standard chart-based install; the entry point that ships with sensible defaults.
• ScaledObject CRD. Defines what scales (deployment) and based on what signal (Kafka lag, queue depth); declarative configuration.
• Cooldown period. Per-ScaledObject cooldown prevents thrashing on transient signal dips.
• Min and max replicas plus documented trigger. Explicit bounds avoid surprise scale-to-zero or runaway scale; per-workload scaling source committed to the repo.

Why this compounds

KEDA mastery compounds across event-driven workloads. Each ScaledObject the team writes becomes a template the next workload inherits; each scaling event teaches workload patterns; the team builds intuition for matching scaling signals to workload shape.

• Cost efficiency. Scale to zero on idle; the workload consumes capacity only when there is real work to do.
• Scaling fit. Match scaling to actual demand signal; the workload tracks the queue rather than tracking CPU as a proxy.
• Reusable patterns. Standard ScaledObject templates capture conventions; new workloads inherit the patterns.
• Institutional knowledge. Each scaling event teaches workload patterns; the team builds vocabulary for event-driven scaling.

The first KEDA setup is an infrastructure investment that pays off across years. Nova AI Ops integrates with scaling telemetry, surfaces ScaledObject patterns, and supports the team’s auto-scaling discipline.