Traditional CPU/RAM metrics miss engine queues and frozen users. Scrape vLLM-native /metrics for TTFT, ITL, KV cache usage, queue depth. Gateway already has prometheus annotations (deployment.yaml:30-32) but vLLM pods don't. Add ServiceMonitor for llm-serving namespace. Files: k8s/apps/llm-serving/, k8s/infra/monitoring/
No dependencies set.
The note is not visible to the blocked user.
Traditional CPU/RAM metrics miss engine queues and frozen users. Scrape vLLM-native /metrics for TTFT, ITL, KV cache usage, queue depth. Gateway already has prometheus annotations (deployment.yaml:30-32) but vLLM pods don't. Add ServiceMonitor for llm-serving namespace. Files: k8s/apps/llm-serving/, k8s/infra/monitoring/