Skip to content
Tech Interview Prep
Technical interview guide

Translating Requirements into Architecture

Turning a specific customer's stated needs and constraints into a concrete, deployable system design.

Read
42 min
Practice MCQs
25
Interview Q&A
25
Edition
v4
Editorial status
Reviewed

Scope: Azure and AWS Well-Architected guidance, Google Cloud Architecture Framework, SEI quality-attribute methods, and C4 model guidance current 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-3A design doc for a checkout rewrite splits its entries two ways. Requirements holds whatever a duty, a committed date or a measurement has already settled, and each entry carries an acceptance test. Assumptions holds what nothing has settled yet, and each entry carries an owner and a validation date. Which entry belongs in Assumptions?
MCQ-7A repo stores one ADR per numbered Markdown file, each with a `Status:` line. ADR-0014 (Status: accepted) chose PostgreSQL for the ledger in March. In August a load test at 5,000 writes/s showed p99 write latency of 210 ms against a 50 ms budget, so the team moved the ledger to CockroachDB. The repo's CI check fails any build where two ADRs for the same component are both `accepted`; the auditor must be able to retrieve the March rationale; and implementers need exactly one accepted ADR for the ledger. Which file set satisfies all three?
MCQ-10A team's context view for 'Billing' shows Billing, its users, and four external systems. A new engineer asks why 'Ledger' is missing. Billing calls Ledger on every invoice; the same team builds and deploys Ledger from the same repo, and it runs as its own process in the same cluster. Where does Ledger belong in the model?
MCQ-6A team is about to add a service that must sustain 20,000 requests/sec with p99 under 10 ms. They find an accepted ADR from 2023 that says: 'Decision: use REST over HTTP/1.1 for all internal service calls. Context: six-week delivery deadline; team already knew REST. Alternatives: gRPC, GraphQL. Consequences: simpler debugging, higher per-call latency.' The ADR has no other entries. What does the ADR's content most directly support?
MCQ-2A dispatch team selected Product A for their batch job system after a two-week pilot showed 40,000 msg/s at p99 8 ms, beating Product B's 22,000 msg/s at p99 11 ms. Their measured peak is 3,000 msg/s. Two weeks after launch, batch workers that stay offline for more than 4 hours lose messages. Product A holds messages for offline consumers for 4 hours; Product B holds them for 7 days. The workers can be offline for up to 12 hours. What is the most likely explanation for the message loss?
MCQ-1A checkout page must render within 2 s at p99 on mobile, at peak load of 500 rps. It makes three sequential API calls: inventory (p99 400 ms), pricing (p99 300 ms), and charge. The charge call's p99 includes a 200 ms auth hop that the team measures separately and leaves out of its own dashboards. The team is replacing 'the charge API must be fast' with a measurable requirement. Which one is correct?
MCQ-4A team drafts an availability scenario for checkout: "When one of the two availability zones fails during peak shopping, checkout still returns successful responses to 99.9% of requests at 12,000 requests/s, with p99 under 250 ms across the incident." Load tests show a single zone serves 7,000 requests/s and the load balancer returns 503 once a zone is at capacity. At peak, 12,000 requests/s arrive and are spread evenly. A game day takes one zone down for ten minutes at peak. Which statement about that run is correct?
MCQ-11A per-service latency dashboard sums to 230 ms (gateway 20 ms, orders 50 ms, payments 70 ms, ledger 90 ms), but a traced checkout request at p99 takes 1800 ms. The trace shows payments calling the ledger, getting a 503, retrying twice at 90 ms each, then publishing `payment.captured`; orders consumes the event 1300 ms later and writes a second ledger entry at 90 ms. Which conclusion does the trace support?
MCQ-17A checkout request must meet a 400 ms p99 budget. The critical path is edge, API gateway, order service, a third-party payment authorization, a third-party fraud check, and the orders database. Measured p50/p99: edge 10/25, gateway 8/20, order service 20/70, payment 85/150, fraud check 30/60, database 12/40. The payment provider offers no p99 SLA, only a 300 ms timeout; the fraud check's provider commits to a 70 ms p99. Which allocation should the team adopt?
MCQ-19A payments ledger runs in a single region. Its recovery plan is a nightly full backup plus asynchronous replication to a standby region. In the last drill, restoring from backup took 3 h 40 m, and replication lag was 1.5 s at the median and 95 s at p99 during the 02:00–03:00 settlement batch. The business has signed off an RTO of 15 minutes and an RPO of 30 seconds for this ledger. Which conclusion about the current design is correct?
MCQ-9A payments service keeps a bidirectional trace matrix: every requirement row lists the design decisions that satisfy it and the tests that verify it, and the decision rows list the requirements they serve. The matrix below is current as of this morning. | ID | Requirement | Design decisions | Verifying tests | | --- | --- | --- | --- | | R-1 | Idempotent retries keyed by request ID | DD-3 | T-11, T-12 | | R-2 | Retry window at most 24 hours | DD-3, DD-9 | T-13 | | R-3 | Every retry attempt logged | DD-9 | T-14 | | R-4 | Settlement file encrypted at rest | DD-14 | T-15 | | R-5 | Duplicate settlement suppressed | DD-3 | none | | R-6 | Merchant-visible retry counters | DD-21 | T-16 | DD-3, the retry ledger key, is being revised and the change is ready to ship. Which tests must be re-run to re-establish verification for the requirements DD-3 satisfies?
MCQ-18Checkout Service adds a card_bin field to the PaymentAuthorized event it publishes and logs the full event at INFO. Payment Service and Ledger Consumer read that event, and every service writes logs through a shared Logging Library. The team's traceability tool links each requirement to exactly one component, so its dashboard reports 412 of 412 requirements linked, 100%. The change-impact query -- which requirements are linked to component X? -- run against Checkout Service returned exactly one requirement, PAY-09, because the other five requirements the post-incident review found were in scope for this change are filed elsewhere: PAY-14 (card data must not appear in application logs) sits on the Logging Library, and the schema-version, BIN-masking and key-availability requirements sit on the Schema Registry and the Key Service. Which change to the requirement links makes that same query, run against Checkout Service, return all six?
MCQ-12A team is choosing how to run a new event pipeline. They score three options on a 1–5 scale (5 best) across three weighted criteria: throughput (weight 0.4), ops cost (0.3), migration effort (0.3). The managed service scores 5, 4, 3; self-hosting scores 4, 3, 4; the current pipeline scores 3, 5, 5. The team recommends the managed service. What should the team do?
MCQ-13A sourcing team scores two vendors on three criteria in a weighted matrix: Cost (weight 0.5), Performance (0.3), Support (0.2), each scored 1–10. Vendor A scores 9/5/5 and Vendor B scores 6/9/6, so A leads B by 0.1. The team reruns the matrix with Cost at 0.4 and Performance at 0.4. What should the team conclude from the rerun?
MCQ-14A payments team must choose between calling a vendor's fraud-scoring API synchronously and building an async local model. The SLO is p99 < 150 ms at 500 rps. The vendor's docs claim p99 80 ms at 500 rps but say nothing about behavior when the API returns 429s. The team has 10 days before committing to a 6-month build. Which PoC plan best retires the decision-critical unknown?
MCQ-15A team is deciding whether to replace their session cache. The vendor's benchmark runs on an 8-vCPU node with 64 GB RAM, 1 KB values, a 4 GB dataset, 100% reads and 50 clients, reporting 480k ops/s mean and 0.9 ms p99. Production runs on 64-vCPU nodes with 64 GB RAM, serves 40 KB session blobs from a 300 GB working set, is 25% writes, and must hold p99 under 5 ms at 2,000 concurrent clients. What should the team do?
MCQ-8A platform team's architecture review board meets two hours a week. Its charter: escalate a decision only if it is a one-way door. Reversing it would cost more than 40 engineer-days, or it would break a commitment already made to a party outside the team. Five decisions are pending. Which one should be escalated?
MCQ-5A payments platform runs 40 business flows under one platform-wide 99.9% monthly availability SLO and one shared error budget. Last quarter the recommendation-widget flow, a non-critical sidebar, burned 62% of that budget on a flaky downstream dependency, and the resulting two-week release freeze also blocked the fix for a checkout-flow bug that had already caused three customer-visible payment failures. Which change addresses the root cause?
MCQ-16A payments service must meet a contractual p99 write latency of ≤ 80 ms. A compliance mandate requires every write to be synchronously replicated to a second region before the client gets a 200, with a regulatory deadline in three weeks. Measured: sync replication adds 95 ms, pushing p99 to 175 ms. Async replication holds p99 at 80 ms but leaves a replication-lag window of up to 200 ms. The VP of Payments owns the latency SLO and has pre-approved a written waiver for any p99 increase up to 100 ms. The Head of Compliance owns the mandate and has pre-approved a written waiver for any replication-lag window up to 150 ms. No budget exists this quarter for a faster inter-region link. What should the tech lead do before the release?
MCQ-21A payments service passes its steady-state load test (p99 40 ms at 3,000 rps) and its architecture requirements cover only that request path. In six months of production it has had four multi-hour outages, each beginning with a routine change: a config rollout, a schema migration, a certificate rotation, a node-pool upgrade. Post-incident reviews found no way to see which build each instance is running, no way to drain a node, and no rollback for the migration. Which change to the architecture requirements does this evidence support?
MCQ-20A team is three weeks into replacing a nine-year-old billing monolith with services. Their assessment catalogued 42 HTTP endpoints, mapped the target domain model, and picked Kafka for events. In the first cutover rehearsal the new service passed all 42 endpoint tests, but finance found month-end invoices wrong for 1,200 accounts: the legacy nightly reconciliation job reads the `invoices` table directly with a read-only account, and the new service no longer writes the rows that job expects. Which gap in the assessment does this failure expose?
MCQ-24ADR-014 chose a single Postgres primary with two read replicas for the orders service, on the recorded assumption that sustained writes stay under 200/s. The ADR's revisit trigger was 'revisit if sustained writes exceed 150/s for three days'. A quarter later, telemetry shows 180/s sustained writes for the past week, and the orders SLO (p99 < 100 ms) is still met at 62 ms. What should the team do?
MCQ-22An order service and a ledger service each own a separate database and are deployed independently. A new requirement says a captured payment and its ledger entry must commit or fail together, and auditors must read both in a single query. The ledger service is also unavailable for 20 minutes each night during its batch window, and captures must still succeed then. Which change does that requirement force?
MCQ-25A payments platform runs 40 services, each with its own datastore and its own authorization code. A new regulation requires that every read of cardholder data be tied to an approved business purpose and be recorded in one auditable log. Which design change does the requirement force?
MCQ-23A payments team ships in two weeks a release that also carries the fix for an actively exploited auth bypass, so the release cannot slip. One requirement in it, moving the legacy batch path onto tokenized card storage, needs six more weeks of work. The batch path carries 3% of card volume, and the card-brand audit is in November, four months out. The product owner wants to ship on time and defer the tokenization requirement. Which deferral is safe?