Monitor GenAI applications beyond golden signals | ODSP907
GenAI applications require extending traditional golden signals because they are non-deterministic, variably costed, newly attackable, and subjectively judged.
“What makes Genai applications fundamentally different from the traditional software we've been monitoring for decades? There are four key shifts.”
A Microsoft Build session argues that traditional monitoring's golden signals (latency, errors, traffic) must be adapted for GenAI's non-determinism, variable cost, new attack vectors, and subjective quality. It offers practical guidance like instrumenting latency per pipeline stage (RAG retrieval, LLM call) and segmenting errors by model and user. Useful engineering practice but a general best-practices talk rather than a major industry signal.