Cost-aware Kubernetes autoscaling with KEDA and Karpenter

Autoscaling in Kubernetes can often feel like a balancing act: HPA does not scale fast enough, cluster autoscaler lags behind, and costs spike when workloads surge. The platform team spent the last year building a unified autoscaling strategy using KEDA for flexible, event-driven workload scaling and Karpenter for fast, cost-aware node provisioning. The results fundamentally changed how the team operates.

In this talk, Ji will discuss:
• Why combining KEDA and Karpenter gave the team predictable scaling across all layers
• How the team expanded beyond basic KEDA scalers to support more event sources
• The tuning, metrics, and thresholds that worked for real workloads
• The unexpected behaviors and anti-patterns the team experienced along the journey
• The annual engineering hours saved after adopting Karpenter
• Cost optimization wins achieved through consolidation, right-sizing, and faster scale-down
• Practical guidance for platform teams adopting both tools together

Whether attendees are running small clusters or scaling a multi-team platform, this session offers a realistic, experience-backed guide for building an autoscaling setup that reacts quickly, stays stable, and keeps cloud costs under control.