Skip to main content

Quiz Calibration

Attempt the quiz before reading this page. Preserve the original answer, correction and help received. These explanations calibrate retrieval; successful recall alone does not complete the module or its practical evidence gates. For context such as shared schemas, use the question page.

Question 1: Requirements Discipline​

You are asked to "design a notifications service." What are the first three things you must explicitly ask before any design?

Calibration: (1) Who sends, who receives, and through what channels (functional scope); (2) scale and latency targets: DAU, notifications/user/day, P99 delivery latency, durability (non-functional with numbers); (3) constraints: mobile push vs email vs SMS, regulatory requirements, existing infrastructure.

Question 2: Estimation Drill -- Feed Writes​

A service has 100 M DAU; each user writes 5 items per day; each item is 500 bytes. Estimate (a) writes per second average, (b) raw storage per day, (c) raw storage per year.

Calibration: (a) 100 M × 5 / 86,400 ≈ 5,800 writes/sec average (~5.8 K/s; peak at 3× ≈ 17 K/s).
(b) 100 M × 5 × 500 B = 250 GB/day.
(c) 250 GB × 365 ≈ 91 TB/year raw (before replication/indexes); with 3× replication and ~30% index overhead, call it ~360 TB/year.

Question 3: Estimation Drill -- Cache Sizing​

Your read workload is 80/20 skewed. You want 95% cache hit rate on the last 24 hours of content, where 24 hours = 1 TB raw of items. Roughly how much cache memory do you need?

Calibration: The 80/20 rule says ~20% of items serve ~80% of requests; to get to 95%, you cache somewhat more. A reasonable heuristic is 25-35% of the hot period, so ~250-350 GB of distributed cache. State the heuristic; the exact number is less important than showing you can derive it.

Question 4: Estimation Drill -- Latency Budget​

P99 target 200 ms. Client to edge: ~30 ms. Cross-continent round trip: ~150 ms. How many cross-continent round trips can the request path afford?

Calibration: Zero on the hot path. Even one cross-continent RTT eats 150 ms and leaves only 20 ms for everything else including client-to-edge. Cross-continent calls must be avoided on the user-visible request path; use async replication, regional caches, or region-local writes with eventual cross-region sync.

Question 5: Pick the Tradeoff -- Storage​

You are designing a time-series metrics pipeline, 1 M data points per second, 100 B each, with mostly-append writes and aggregate reads over recent windows. What storage would you choose and what would you reject?

Calibration: Choose a time-series or columnar store (InfluxDB, ClickHouse, TimescaleDB, or Prometheus for short-retention). Reject OLTP SQL because the write volume (~100 MB/s) will saturate B-tree indexes and reads will sweep enormous row ranges; reject a generic document store because aggregate reads over billions of tiny points need columnar layout to be fast. Cost: less flexibility in ad-hoc schema evolution; mitigated because the schema is narrow.

Question 6: Pick the Tradeoff -- Consistency Model​

Two users race to reserve the last seat on a flight. What concurrency mechanism do you pick, what do you reject, and why?

Calibration: Pick optimistic concurrency with a version column on the seat row: UPDATE seats SET held_by=?, version=version+1 WHERE seat_id=? AND version=?. Reject last-writer-wins (would double-allocate) and reject pessimistic locking (would block and create a stampede). Cost of OCC: one retry on conflict for the loser; in a 10-seat race among 10K users, most requests fail fast with "taken", which is the correct UX anyway.

Question 7: Partitioning Key​

A social feed stores one row per (user, post) in a user_timeline table. Should you partition by user_id or by author_id? Why?

Calibration: Partition by user_id. Reads are always "give me user X's timeline", which must hit one partition. author_id would concentrate load on celebrity authors (hot partitions) and scatter each user's timeline read across many partitions, which is the opposite of what you want.

Question 8: Cache Placement Critique​

A junior engineer proposes putting a cache in front of a billing ledger to "speed up reads". Why is this a correctness risk, and under what condition could it be acceptable?

Calibration: The ledger is a source of truth for money; a cache that can return stale balances can cause a user to overdraft or double-spend if downstream logic acts on the cached view. It is acceptable only if (a) the cache is explicitly a read-only summary for display purposes, (b) all money-moving operations read directly from the ledger, and (c) the UX makes clear which numbers are cached vs authoritative.

Question 9: Bottleneck vs SPOF​

A centralized ID-generation service runs on a single instance. It serves 5 K QPS at 10% CPU. Is it a bottleneck? Is it a SPOF? What do you do?

Calibration: It is not currently a bottleneck (plenty of CPU). It is a SPOF because its death blocks all new writes system-wide. Remediation: either make it highly available (active-passive with fast failover) or redesign as distributed range-allocation so that instance death has bounded blast radius (each service pre-allocates ID ranges and can operate from its range for minutes without the allocator).

Question 10: 10x Scale Reasoning​

Your system handles 10 K QPS today with a single Redis cache (50 GB RAM, 80% hit rate). At 10× traffic (100 K QPS, same workload shape), what breaks first and what do you change?

Calibration: The cache breaks first: with 100 K QPS and the same hit rate the origin sees 20 K QPS, up from 2 K, which may saturate the DB; and Redis throughput per node caps well below 100 K QPS for non-trivial commands. Remediation: shard Redis across 3-6 nodes with consistent hashing; grow hot set to 100-200 GB; add a short-TTL edge or per-instance cache to absorb very-hot keys; re-check DB with the new cache-miss rate.

Question 11: Failure Walk -- AZ Outage​

Your feed service has app servers and a cache in every AZ of one region, plus a primary DB in AZ-1 with a synchronous replica in AZ-2. AZ-1 goes dark. What happens, and what is your RTO/RPO?

Calibration: App servers in AZ-1 and their local cache die (load balancer ejects them); traffic shifts to AZ-2/AZ-3 stateless tiers. The DB primary in AZ-1 is lost; AZ-2's synchronous replica is promoted. RPO ≈ 0 (synchronous). RTO ≈ 30-120 seconds (health-check-driven failover and promotion). Warm-cache tiers in AZ-2/AZ-3 absorb traffic; origin sees a spike until cold-cache users in AZ-1 are rerouted.

Question 12: Pick the Tradeoff -- Fan-out​

For a chat app with groups of 1-100,000 members, compare "fan-out on write to every recipient's mailbox" vs "fan-out on read by joining on the group". Which do you pick for which group sizes?

Calibration: For small groups (say < 1000), fan-out on write: the per-message write amplification is bounded, and reads become cheap keyed lookups. For very large groups ("channels"), fan-out on read: writing 100 K rows per message is wasteful, and members who never read never pay the cost. The hybrid is to threshold on group size, pay fan-out write for small groups, and use a read-time view for large channels. Cost of the hybrid: two code paths; mitigated by a shared abstraction.

Question 13: Trade-off Articulation​

Write a single-sentence trade-off statement in the form "I chose A over B because C, accepting cost D" for the decision "use cache-aside with explicit invalidation".

Calibration: "I chose cache-aside with explicit invalidation over write-through caching for the profile store because write-through couples every profile write to the cache tier's availability and we cannot afford that coupling, accepting a brief staleness window after a write until the next reader repopulates the cache."

Question 14: Four-Phase Discipline​

You are 22 minutes into a 45-minute interview and still in phase 2 (high-level diagram). What do you do?

Calibration: Stop drawing. Say out loud: "Let me close out high-level and move to deep-dive; I'll come back to any gaps if time allows." Name the diagram as "committed as-is", pick a component to deep-dive (or ask the interviewer), and allocate no more than 15 minutes to phase 3. Protect the last 3-5 minutes for wrap-up regardless. Running over phase 2 into phase 3 is a known failure mode; cutting your losses is the senior move.

Question 15: Design Doc Minimum​

Name three sections of a design doc that, if missing, cause a senior reviewer to reject the doc.

Calibration: (1) Requirements with numbers (non-functional targets quantified). (2) Scale and failure analysis (10×, AZ outage, per-box failure walk). (3) Trade-offs (explicit decisions with rejected alternatives and costs named). Losing any of the three indicates the author did not understand the shape of the problem.

Prior Module Question 1 (S7 Architecture & DDD)​

Why is a bounded context a design tool rather than just a naming convention?

Calibration: A bounded context defines where a set of domain terms has a single, consistent meaning and where transactional integrity holds. Crossing it requires explicit translation (contracts, anti-corruption layers). Without bounded contexts, shared models leak coupling and force unrelated teams to agree on vocabulary.

Prior Module Question 2 (S7 API Design)​

What is the difference between an API's contract and its implementation, and which one are consumers coupled to?

Calibration: The contract is what consumers can observe (request/response shapes, status codes, idempotency, ordering guarantees); the implementation is how the service achieves it. Consumers are coupled to the contract. Changing implementation freely is safe; changing the contract silently breaks consumers.

Prior Module Question 3 (S6 Distributed Systems)​

State the CAP theorem in one sentence and name which property modern systems typically preserve during a network partition.

Calibration: CAP says a distributed system can only provide two of consistency, availability, and partition tolerance at a time; during a real network partition (which is unavoidable), the choice is between consistency and availability. Most user-facing systems choose availability and recover consistency afterward; systems handling money or identity typically choose consistency and accept unavailability.

Prior Module Question 4 (S6 Databases)​

Why is a secondary index on a write-heavy table expensive, and what is the typical diagnostic when indexes are overused?

Calibration: Each secondary index must be updated on every insert/update/delete of the indexed columns, turning one write into N+1 writes; it also occupies storage and buffer-pool space. The diagnostic is rising write latency and IO utilization correlated with recent index additions.

Prior Module Question 5 (S7 Architecture)​

What is the difference between a quality attribute (architecture characteristic) and a functional requirement, and why do architects pay more attention to the former?

Calibration: A functional requirement defines what the system does; a quality attribute (scalability, availability, latency, security) defines how well. Two very different architectures can both meet the same functional requirements; the quality attributes are what separate them. Architects choose shape primarily on quality attributes because they are what the architecture is trying to deliver.