KubeCon + CloudNativeCon Europe 2025

Balancing resource provision for LLM workloads is critical for maintaining both cost efficiency and service quality. Kubernetes’s Horizontal Autoscaling offers a cloud-native capability to address these challenges, relying on the metrics to make the autoscaling decisions. However, the efficiency of metrics collection impacts how quickly and accurately Autoscaler responds to the LLM workload demands. This session explores strategies to enhance metrics collection for autoscaling LLM workloads with:
1. The fundamentals of how horizontal autoscaling works in Kubernetes
2. The unique challenges of autoscaling LLM workloads
3. A comparison of existing Kubernetes autoscaling solution for custom metrics with their pros and cons
4. How optimizing metrics collection through push-based approaches can improve scaling responsiveness.
It will demonstrate an integrated solution using KServe, OpenTelemetry collector and KEDA to showcase how they can be leveraged to optimize LLM workload autoscaling.