Skip to content
Tech Interview Prep
Technical interview guide

Technical Stakeholder Presentations

Presenting an architecture to both technical and non-technical stakeholders in the same room, credibly.

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

Scope: Google Technical Writing, W3C WAI, Azure Well-Architected, and AWS Well-Architected guidance current 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-21Two teams run the same weekly 30-minute architecture decision meeting with the same 8 attendees. Both send the full 20-page design doc 24 hours ahead, and every attendee confirms they read it. Team A's doc opens with a one-page brief that states the decision and what each option costs; Team B's recommendation sits on page 17. Over 10 weeks, Team A reaches a decision in 9 of 10 meetings (median 26 minutes); Team B reaches one in 4 of 10 (median 48 minutes), and in its 6 overruns the author spends the first 20 minutes walking pages 1–12 in order. Which change to Team B's pre-read most directly closes the gap?
MCQ-4A payments team took a 22-page design doc into review with three staff engineers, a security reviewer, and a VP who had 10 minutes before her next meeting. The VP approved after page 1; the threat model on pages 14-22 held an open question about key rotation, and three weeks into implementation the team had to take the approval back. The doc's content and the attendee list stay the same for the re-run. Which structure should the doc take this time?
MCQ-3Your team has one 30-minute slot before the investment committee to get a two-quarter rewrite of the checkout pipeline funded. The CTO, who signs the budget, is there for the first 10 minutes and then leaves for a board call; the other five attendees - two product managers, a finance analyst and the two engineers who wrote the proposal - stay for the whole slot, and none of the four non-authors has opened the 38-page design doc. The rewrite's figures are p99 checkout latency 1.2 s down to 300 ms and on-call pages 9 a week down to about 2; the cheaper incremental fix was killed because the vendor library it leans on reaches end-of-life in 7 months. Committee rules defer any proposal that does not name a decision to the next quarterly cycle. Which opening do you lead with?
MCQ-1You have a 30-minute slot at the quarterly architecture review to get approval for a $400k, two-quarter migration off a self-managed Kafka cluster. The committee is the VP of Engineering (who holds the budget), two directors, a security lead, and a finance partner. The current deck is 42 slides: 12 on the cluster's history, 15 on migration options, 8 on benchmark results, 5 on the recommendation, 2 on next steps. The last two versions of this deck ended with 'let's revisit next quarter' and no decision; the migration has now slipped two quarters, and the self-managed cluster runs $18k/month more than the managed alternative. Which change to the deck is most likely to produce a decision in this slot?
MCQ-2Thirty minutes are on the calendar to decide whether to shard the billing database. In the room: the VP of Engineering, who approves the budget and has a hard stop at minute 20; the payments tech lead, who owns the client retry code; two on-call SREs who will operate the shards; and a PM who owns the customer cutover date. In the pre-read thread, the tech lead asked who rewrites the retry logic and the SREs asked how failover works. The VP has said she will approve if the migration's cost fits the existing on-call budget, and the retry-logic answer is what sets that cost. Which agenda gets a decision in the room?
MCQ-18A 45-minute architecture review has 9 attendees and six contract sections budgeted at 6 minutes each, with 9 minutes of buffer. At minute 22, section 3 is still open when a staff engineer raises a data-residency concern about a service that ships two quarters from now. The internal API contract the meeting was convened to approve does not touch that data path, the data-platform lead who owns the concern is not in the room, and working it through would take about 25 minutes. The sprint depends on today's contract approval. What should the facilitator do?
MCQ-11A staff engineer's service-split proposal goes to the architecture review board in five days. The board has six voting members, and its charter requires the decision and any conditions to be recorded in the meeting. In her last three reviews the proposal passed, but each time the board added a condition costing 4–8 engineer-days of rework; in two of the three, the condition came from a member who had raised the same concern with her privately the week before. Which plan for the five days best reduces the rework?
MCQ-24At minute 52 of a 60-minute design review, the principal engineer has held the floor for 19 of the 25 Q&A minutes on partition-count math. The review must decide whether to adopt the vendor's event bus; its charter requires the security lead and the SRE to record a position before the decision, the vendor's pricing hold expires at the end of the meeting, and the SRE has a hard stop at the hour. Neither the security lead nor the SRE has spoken. What should the facilitator do?
MCQ-13A payments team must choose a replication mode for its ledger database before Friday's launch. Load tests have settled the technical facts: synchronous replication raises p99 write latency from 41 ms to 49 ms and loses no committed writes; asynchronous holds p99 at 41 ms but can lose up to 15 minutes of writes if a region fails. Both modes pass every correctness and capacity test. The SRE lead wants synchronous; the product owner wants asynchronous to protect the launch latency SLO. The payments VP owns the ledger's data-loss exposure and signs off on launch risk. What should the architect do?
MCQ-16A payments company's nightly settlement batch runs in a 00:00–06:00 window and currently finishes at 05:00. Runtime scales linearly with transaction volume, which grows 10% per quarter. Parallelizing the batch costs $400k. The CTO has one slide for the CFO's budget review. Which statement should the CTO put on the slide?
MCQ-17A design review tomorrow covers a 40-service dependency graph with 118 edges. Two of the eight attendees use screen readers, and four join on 13-inch laptops. The graph is currently a 4000×3000 PNG pasted into the design doc. What should you post before the review?
MCQ-15A storage review is 20 minutes from approving a 4-node plan for a 120 TiB working set, sized from a 3.0 compression ratio measured on a 200 GB sample. Each node provides 12 TiB of usable SSD. Mid-review, a full-dataset scan shows the real ratio is 1.8. The deck has been circulated and the sponsor is in the room. What should the review do?
MCQ-8A team has 30 minutes to decide whether the checkout service can move to a new region. Their diagram shows all 18 services in the target state, a legend, and labeled arrows, but it collapses checkout's synchronous call to the payment provider and its asynchronous write to the ledger into one arrow labeled 'payments', and it omits the on-prem fraud service that checkout calls on every order. Which change to the diagram most improves this decision?
MCQ-9Checkout is documented as one diagram: 12 service boxes with arrows, protocol labels, and the owning squad on each box. A post-incident review cannot get two things out of it: which zone the ledger-writer ran in (6 replicas across 3 zones, and only the us-east-1b pair absorbed the retry storm), and the call order for a single timed-out checkout (the gateway called fraud-scorer, got a 504, called fraud-scorer again, then wrote to the ledger). What should the team produce?
MCQ-22A retail platform is six weeks into a fourteen-week migration from a monolith to nine services. The runbook's architecture diagram draws all ten components in the same style, with no state labels. Responding to a queue backlog, an on-call engineer scaled the monolith's order-consumer deployment to zero, believing it had already been replaced; it was still the only writer to the orders table, and 22 minutes of orders were lost. Which single change to the diagram would have prevented that failure?
MCQ-12Your team is sizing a read-through Redis cache in front of Postgres. A 24-hour production sample holds 8,000,000 read requests across 200,000 distinct keys, and 62% of those keys were requested exactly once in the sample. Sorting keys by request count gives: the top 1% of keys (2,000) carry 31% of reads, the top 10% (20,000) carry 70%, the top 25% (50,000) carry 86%, and the top 50% (100,000) carry 96%. The request ranking is stable across the day, so an LRU resident set converges on the most frequently requested keys. You provision 20,000 entries. What hit rate should the capacity plan assume?
MCQ-5You are writing the capacity line for the launch review of a payments API on a 12-node cluster. The review rule is that a committed rate must hold with any single node drained, at p99 at or under 250 ms. Pre-launch results: a four-hour soak with all 12 nodes held 2,400 rps at p99 190 ms; the same soak with one node drained held 1,900 rps at p99 235 ms, and when it stepped to 2,100 rps p99 hit 420 ms and the run was stopped; a ten-minute burst on all 12 nodes reached 2,800 rps at p99 240 ms. The traffic trace projects a 2,000 rps peak hour next quarter. Which line belongs in the capacity section of the review?
MCQ-14A team PoCs a 3-node distributed database with synchronous replication to 2 of 3 nodes. At 12,000 ops/s with 4 KB rows and a 70/30 read/write mix, p99 is 18 ms over 72 hours. Killing one node: throughput holds at 12,000 ops/s, p99 22 ms, no acknowledged write lost. Killing two nodes: the cluster rejects writes with 'no quorum', and no acknowledged write is lost. Which additional claim does the evidence support?
MCQ-7You are presenting an in-place on-disk format migration for approval. A staff engineer asks for p99 write latency at 3× current peak. You have measured 90 ms p99 at current peak on staging and have no 3× figure. The write path's SLO is p99 ≤ 150 ms. The review votes today with 40 minutes left; the next review is in six weeks. The staging rig can replay a captured 3× trace in 8 minutes. The migration rewrites the format in place, so a partial rollout cannot be rolled back. Which action do you take?
MCQ-23An ADR recommends Postgres over DynamoDB for a session store, citing 6 ms p99 writes against DynamoDB's 9 ms. A reviewer finds the Postgres figure was measured at 500 writes/s and the DynamoDB figure at the target 5,000 writes/s, and that Postgres's p99 at 5,000 writes/s is 40 ms. The SLO is 25 ms p99. Which recommendation does the corrected comparison support?
MCQ-6A payments team proposes replacing a single Postgres primary (3k writes/s, 60% CPU) with 16 shards keyed by tenant_id. Their tradeoff slide has three rows: 'Chosen: shard by tenant_id', 'Benefit: 12k writes/s headroom', 'Cost: 3 weeks'. Review stalls on two things the slide cannot answer: the write volume at which the single primary stops being viable, and the residual risk the team is accepting by sharding. Which pair of additions answers both?
MCQ-10A payments API carries a 99.9% monthly availability SLO, 43 minutes of allowed unavailability in a 30-day month, and an error-budget policy that freezes feature launches once the budget is exhausted. Six days into the month the service has logged 30 minutes of full outage. The VP of Product, who owns the launch decision, has 15 minutes and wants to know whether next week's checkout redesign can ship. What should you tell her?
MCQ-20A design review for a new search index set one goal: cut p99 search latency below 100 ms. The baseline was 180 ms p99, measured over a week of production traffic at the production query mix. The review ended with a recorded decision to adopt index design X, which then shipped. Which follow-up result is the strongest evidence the review met its goal?
MCQ-25An ADR written two years ago records the decision to serve all order writes from a single primary Postgres, the constraint 'sustained writes stay under 2,000/s', and the consequence 'revisit if p99 write latency exceeds 50 ms'. Today sustained writes are 6,000/s and p99 write latency is 210 ms. A new engineer proposes sharding the orders table. What does the record establish about the decision?
MCQ-19AR-114 closes with 'all services partition by tenant_id'. The decision is merged into the ADR repo; the action-item table is left empty. The ledger service is greenfield: every service releases through the platform's release pipeline, the ledger design doc records partition_key = account_id in an appendix, a platform engineer signs off its design review, and the release checklist's 'partition key declared' item passes because the shipped config matches the doc. Ledger migrates and releases in week 10. Its cross-shard joins with reporting, which partitions by tenant_id, are exercised only against production data and time out at p99 2.4 s in week 13; the on-call who digs in finds AR-114 in the ADR folder. The architecture review next meets in week 14, and the weekly cross-service sync only covers changes teams raise; ledger raised none. Changing the key after the migration costs a 1.4 TB backfill, about nine hours at 45 MB/s. Which check, added to AR-114's close, would have caught the ledger's divergence before its week-10 migration?