C05 Solutions¶
Reference answers with the reasoning, not just the answer key. Attempt
../exercises/ first — these drills are about producing a design document, and
reading a good one teaches you very little about writing one.
| # | File | The point of the drill |
|---|---|---|
| 1 | 01_design_the_serving_stack_solution.md |
Requirements → pattern → architecture → a latency budget that closes. The graded skill is arithmetic that adds up and a cache key that survives both a rollback and a security review. |
| 2 | 02_rollout_plan_solution.md |
The uncomfortable result: the canary percentages people write down are usually orders of magnitude too small for the guardrail they claim to be watching. A 1% canary needs 427 days to detect a 1% conversion regression. Stages that cannot decide must be labelled as such. |
| 3 | 03_ab_test_critique_solution.md |
Ten defects in one plausible-looking write-up. The headline: the p-value is 1.2 × 10⁻⁶ and the result is still uninterpretable, because a tiny p-value on a corrupted design is a precise measurement of the wrong thing. |
Where reference answers stop being authoritative¶
These are defensible answers, not the only correct ones. Design has legitimate variation, and a different answer with better-stated reasoning beats matching this one. Specifically:
- Drill 1 — packing tenants versus isolating them is a genuine judgement call driven by contractual and risk factors the brief only partly specifies. Both answers can be right; only one of them can be unjustified.
- Drill 2 — a slower, safer rollout is defensible for a marketplace and so is a faster one, if you say what you are accepting. What is not defensible is a stage plan whose guardrails have no statistical power and whose author has not noticed.
- Drill 3 — the defect list is closer to objective. The mechanisms are not matters of taste, and the corrected plan has little room for variation.
What is not negotiable in any answer¶
- Latency budgets sum at p99 per hop and leave headroom.
- The autoscaling signal is not CPU utilization and not raw GPU utilization.
- Cache keys include the model version (and the tenant, where there is one).
- Rollback is faster than the error budget it is protecting, and it has been rehearsed.
- Randomization is on the user, and the sample size was fixed before launch.
- Every design states what it is trading away.