KEDA GPU scaler slashes latency, drops Prometheus chain for autoscaling
The KEDA external scaler reads NVIDIA NVML metrics directly from a DaemonSet, eliminating the dcgm-exporter → Prometheus → KEDA pipeline. It trims metric lag to 2‑4 seconds and lets inference services like vLLM or Triton scale to zero without extra components. Simpler ops, faster reactions for GPU workloads on Kubernetes.