You already know: TTFT, inter-token latency and goodput (Inference fundamentals); load testing and Little's law (Measurement & capacity); model and data monitoring (D11 Model monitoring); the KV-cache cliff (KV cache I and II).
Why this matters for Atlas. Load tests tell you what the system can do; observability tells you what it is doing right now, for real users, and warns you before they notice. LLM serving breaks the usual web-service habits: latency splits into two signals, the leading indicator is a queue and a memory pool rather than CPU, and incidents can start and end between two default metric scrapes. This lesson specialises the golden signals for LLM serving, then gets precise about the instruments: histogram buckets, scrape intervals, dashboards, alerts, SLOs and error budgets.
Where you'll use this. Building the first dashboard for a new model deployment; writing alert rules that page for user pain, not for noise; setting SLOs a team can actually run against.
1. A silent outage: why averages lie
The question. How can a dashboard stay green while users leave?
Start from nothing. Picture a chat product serving 200 requests per second. At 11:47 a few customers upload 12,000-token documents; a batch job then sends 50 long-context requests at once. The queue grows: 2, 12, 45, 120, 340 waiting requests over twenty minutes. p99 time to first token goes from 0.2 s to 22 s. The average latency, meanwhile, rises only from 180 ms to about 420 ms, because most requests are short and still finish quickly. Alerts on the mean never fire; the support queue fills first.