What should you monitor for a model in production?

Operational metrics (latency, throughput, error rate, cost), model-quality metrics (accuracy/precision/recall where labels arrive, or proxy metrics), input data health (drift, missing values, schema changes), and business KPIs the model is meant to move. For LLM apps, also track token usage per feature, hallucination/quality via sampling and evals, and safety flags. The principle: catch problems from data and quality, not just uptime — a model can be 'up' while quietly making bad predictions.