Reference solutions — how a strong candidate runs the round
Stage-by-stage answer keys for the problems you are most likely to face. Each one is written the way a strong candidate would run a 45–60-minute round: what to ask, which numbers to compute and what they decide, the API and data model, the design and its walk-through, the deep dives, the mid-round curveball, failure modes, SLOs and cost, what changes by level, and the four-sentence close.
Run the round first. A solution read before a mock turns the mock into a recital and the score into fiction. The loop that works: run the question cold in the simulator → get the report → read the solution at the stages the report flagged → re-run weeks later.
The five most-asked
| ID | Problem | The insight in one line | Asked most at | Theory it leans on |
|---|---|---|---|---|
| SD-25 | Distributed rate limiter | the limiter’s availability and latency matter more than its accuracy — decide locally, reconcile async | Google, Meta, Amazon, Microsoft | 8, 7 |
| SD-08 | News feed / home timeline | hybrid fan-out with a threshold: push for normal accounts, pull for celebrities | Meta, Google | 5, 6 |
| SD-09 | Chat / messaging | mailbox first, socket second: durable sequenced log + per-device cursors | Meta, Microsoft | 9, 6 |
| SD-15 | Globally distributed KV store | consistency is a per-operation product choice; say what each level costs | Google, Amazon, Microsoft | 4, 3 |
| SD-44 | Payments & double-entry ledger | exactly-once doesn’t exist across the bank boundary — idempotency, pending states, reconciliation | Amazon, fintech, Google | 7 |
Why these five. Between them they are the flagship question at every major company class in the bank, and they cover the five recurring hard problems of system design: fan-out (SD-08), real-time delivery and ordering (SD-09), distributed consistency (SD-15), correctness under failure (SD-44) and hot-path protection (SD-25). Master these and most other questions are recombinations.
How every solution is laid out
| Stage | Clock (60 min) | Section |
|---|---|---|
| 1 | 0:00–0:06 | Clarify — I ask / they answer / why it matters, scope, non-functional numbers |
| 2 | 0:06–0:11 | Estimation — three numbers, each with its consequence |
| 3 | 0:11–0:17 | API and data model |
| 4 | 0:17–0:27 | High-level design (diagram) and the end-to-end walk-through |
| 5 | anywhere | Technology choices — mechanism and trade-off |
| 6 | 0:27–0:45 | Deep dives — the crux first |
| 7 | as given | The mid-round requirement, as a delta |
| 8 | 0:45–0:50 | Failure modes |
| 9 | 0:50–0:54 | SLOs, observability, cost |
| 10 | when asked | The level layer — Staff vs EM vs Director |
| 11 | 0:55–0:57 | The close — four sentences |
| 12 | after | Self-check |
The boxes: Say it in the room — words you can use nearly verbatim · Why? — the reasoning behind a term or decision · The number that matters · Trap — what breaks under a follow-up · Level lens — how the bar shifts by level. Method in full: The interview method.
Coming next
The next solutions in the queue, by how often they’re asked: SD-12 notification platform · SD-20 distributed cache · SD-47 ticket booking under contention · SD-18 file sync · SD-39 ride matching · SD-42 LLM inference serving · SD-55 design review (inverted) · SD-58 SLO design. Contributions welcome — follow this layout (CONTRIBUTING).