2 min lesson
Enumerate the failure branches out loud
Make the call in this situation: "Under a sudden traffic spike that exceeds total provider capacity, which traffic should your admission control shed first and why?" Explain what supports it.
Step 1 of 2
Enumerate the failure branches out loudWhere points are won
Health-based routing sheds traffic off it before the breaker even trips.
Hedge Tab requests to a second provider past the p95 TTFT, capped at one hedge.
Circuit breaker opens, all traffic flows to the fallback ladder.
Quality may drop a notch; availability holds. Say which you're trading.
Admission control sheds low-value Tab first, protects Agent.
Backpressure caps concurrency so you never cascade into the providers.
Prefix/response cache absorbs repeat work.
Route price-sensitive surfaces to cheaper models; flag the UX risk.
Two ways this round goes sideways: designing one uniform pipeline for all three surfaces (it shows you haven't internalized the latency/cost differences) and hand-waving the numbers. If you say "low latency," the follow-up is "how low, measured how and what's your p99?" Have the answer before they ask.
Reserve two minutes for safety. A routing change ships behind a flag, canaries on a small traffic slice per surface and auto-rolls back on an SLO breach. Blast-radius containment - a bad config can't take down all three surfaces at once - is the thing senior reviewers listen for and most candidates forget.
Close by comparing your answer to the worked design in the routing and failover module. The gaps you find are this round's study list.