Events / Portland

AI-Driven Autoscaling: Adding Pods Before the Spike, Not After

Portland Java User Group talk on swapping Kubernetes' reactive pod autoscaler for one that learns from traffic history, cutting cost and latency for JVM services.

Wed, Jul 29, 12:30 AMProFocus Offices · Portland · OR

Why it matters

The talk's own numbers, latency down ~23%, cost down about a fifth, scaling oscillations cut 5x, are exactly the infra-cost tradeoffs tokenmaxxing readers track for any always-on AI or agent service.

The tokenmaxxing angle

Not about LLM tokens directly, but the same optimization problem: predicting load instead of reacting to it. The approach (add capacity ahead of the spike) maps directly onto right-sizing agent infra spend.

From the organizers

Speaker is Shalini Sudarsan, DevOps Engineering Leader at KinderCare Learning Companies; she reports latency down ~23%, cost down about 20%, and oscillations cut from 15/hour to 3/hour.