You've finished the chapter. GPU procurement, geo-routing, SLO engineering and error budgets, security and compliance, zero-downtime deployment, cost estimation and break-evens, four-signal observability, async inference patterns, streaming protocols, and the build-vs-buy decision. This is the final exam — eight questions that span the second half of the chapter. Clear it and the chapter is complete.
Score 80% or better to pass. Every option explains itself — a wrong pick still teaches — and you can retake any time.
01 · The exam
Infrastructure, ops, cost, and the client layer.
GPU procurement tiers, WAN latency and geo-routing, error budgets and SLOs, deployment strategies, cost break-evens, observability golden signals, streaming protocols, and the managed-vs-DIY crossover.
Check your understanding0 / 8
Question 1
You need sustained H100 capacity for a production 70B inference service. What capacity mix minimizes cost?
Question 2
A user in Sydney makes a request to an LLM API hosted entirely in us-east-1. Roughly how much latency does the WAN round-trip add to their TTFT?
Question 3
Your SLO is 99.5% availability. Your error budget for the month is 3.6 hours of downtime. Halfway through the month, you've consumed 3 hours. What should you do?
Question 4
Which deployment strategy routes a small fraction of live traffic to the new version before promoting it, allowing you to catch regressions on real load?
Question 5
At what rough steady-state RPS does DIY GPU serving (4×H100 reserved at $4/GPU/h, ~50 RPS/set) become cheaper than a managed API (e.g. Together AI Llama-3 70B at $0.88/M tokens, 200 tokens/request)?
Question 6
The four golden signals for an inference service are:
Question 7
You're building a real-time voice agent: the client sends audio frames to the server, and the server streams TTS audio back. Which streaming protocol fits?
Question 8
An AI startup at Series A, 8 engineers, running ~15 RPS steady-state on GPT-4o for their core product — should they self-host?