Skip to content
Tech Interview Prep
Technical interview guide

Proof-of-Concept Design

Scoping a POC narrowly enough to prove the riskiest assumption, without building a miniature version of the whole system.

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

Scope: AWS and Azure architecture and experimentation guidance plus Google Cloud load-testing and deployment guidance current 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-9A payments API serves 100,000 requests an hour. Before a release, mean latency is 50 ms and p99.9 is 500 ms. After the release, the mean is 52.5 ms and p99.9 is 3,000 ms. The release gate blocks only when the mean exceeds 60 ms, so the release shipped and the hour was marked healthy. Which statement about that hour is correct?
MCQ-18A team benchmarks a JSON parser on a dedicated 8-core host, parsing the same 50 MB file in each of 20 runs. Nineteen runs finish between 1.0 and 1.2 ms (median 1.1 ms); one run takes 47 ms. The harness log for that run shows a scheduled OS update pinned one core for its whole duration. The team must publish a single latency figure for a comparison table. What should they publish?
MCQ-8A team benchmarks two candidate RPC frameworks for an ingest service whose requests are 1 KB. The client and server run on separate hosts joined by a 1 GbE link (125 MB/s line rate), and the harness reports client-side end-to-end throughput: framework A 90k req/s, framework B 240k req/s. Which conclusion do these numbers support?
MCQ-7A team is replacing its JSON parser. On the same 1 GB corpus and hardware, the current parser sustains 120 MB/s with p99 8 ms and 200 MB peak RSS per worker. The candidate sustains 150 MB/s with p99 12 ms and 400 MB peak RSS. The service SLO is p99 < 10 ms and the per-worker memory budget is 300 MB. What does the baseline comparison establish for this decision?
MCQ-1A team has one week before a build-vs-buy decision on a managed vector index. The vendor's benchmark claims 40 ms p99 at 200 QPS on 1M vectors, measured with tuned HNSW parameters (m=48, efSearch=256). The team's two-week PoC ran the same 1M-vector corpus at 200 QPS with the client library's defaults (m=16, efSearch=64) and measured 95 ms p99 with recall@10 of 0.91; the current system gives 120 ms p99 at recall@10 of 0.98. Which claim can the PoC's results support in the decision memo?
MCQ-13A vendor runs a PoC for their analytics engine on a single 512 GB node, loading a 40 GB dataset that stays resident in memory, using their own harness. They report 1.8 s for the buyer's ten heaviest queries. The buyer's production dataset is 60 TB across 12 nodes, and those ten queries take 90 s today. The harness log shows each query is executed twice and only the second run is reported. What does the 1.8 s figure support?
MCQ-24A team sized a service from a PoC. The service runs 4 instances, each with a 10-thread request pool, in front of a Postgres 15 instance. The PoC drove it with a closed-loop generator running 40 concurrent clients, replaying a captured production request mix against a full-size copy of the 90 GB production dataset. It sustained 1,000 req/s at p50 40 ms and p99 60 ms for six hours, with app CPU at 35% and Postgres at 40 of 100 connections. Production runs the same 4 instances and the same Postgres class, but the load balancer sends an open-loop 1,100 req/s. For the first few minutes p99 holds near 60 ms, then it climbs past 900 ms and completed throughput settles near 300 req/s. During the collapse app CPU is 35%, Postgres connections peak at 40 of 100, and the DB host NIC runs at 12% of 10 Gbps. What best explains the collapse?
MCQ-10A team runs a PoC of a write-through cache before a production rollout. The PoC uses one 8 GB node holding a 6 GB dataset, uniform key access, 300 writes/s, no deletes, and one region; it measures a 4 ms p99 read latency. Production will use 12 nodes over a 40 TB dataset with Zipfian access (a hot 2% of keys), 3,600 writes/s, 25% deletes, and three regions. The reviewer must decide whether the 4 ms figure transfers. What should the PoC report state?
MCQ-2An event-ingest service must sustain 40k writes/s at peak with p99 write latency under 20 ms. Postgres, which the team already operates, runs the service today at 60k writes/s with p99 at 14 ms. A wide-column store the team has never run would cut storage cost by about 40%. The vendor's benchmark puts that store at 50k writes/s, but it was run with a uniform random key generator; in production the top 1% of devices produce 30% of all writes. The choice is due in six weeks. What should the team do first?
MCQ-3A payments team has four weeks and one staging cluster to decide whether to adopt a managed log service. Their pipeline pins each account to a single partition, and last year a broker failure silently reordered messages and corrupted balances. Which PoC question should they commit to?
MCQ-15A team is running a 10-day PoC of a vendor's enrichment API. The contract caps the account at 50 concurrent connections and the vendor will not raise it; the team must keep all 12 worker processes. At 8 connections per worker the job runs cleanly for about 90 seconds, until every worker has finished ramping its pool, and then nearly every request comes back HTTP 429. The client already retries with exponential backoff and jitter, and the failure rate does not fall. A benchmark against an uncapped account showed the job finishes inside the 10-day window only while at least 40 connections are in flight. What should each worker's pool size be?
MCQ-17A service calls a dependency through a client that retries twice per request (so up to 3 attempts per request). The circuit breaker opens after 5 failed attempts within a 10 s window. During a total dependency outage, the team wants to verify that the breaker caps retry amplification. Which failure test will show that cap?
MCQ-22A payments team spent six weeks evaluating a fraud-scoring service in a dedicated cloud account. On Friday at 18:00 they ran their teardown script, which deletes the compute instances, the load balancer and the database; it is idempotent, ran once, and exited 0. The console then showed all three gone. On Monday at 09:00 all three existed again and the account had resumed billing. No engineer touched the account over the weekend. Which leftover from the PoC explains the reappearing resources?
MCQ-19A team runs a three-week PoC of a candidate time-series database. The PoC replays 30 days of recorded production metrics (4.1B samples) into a five-node cluster in a single-AZ deployment, with authentication disabled and no backup configured. Exit criteria: ingest sustains 20k samples/s, p99 latency for the ten dashboard queries stays under 120 ms, and the cluster recovers within 90 s after one node is killed. All three pass. Which claim does the PoC not support?
MCQ-14A two-week spike for a fraud-scoring service shipped with a hardcoded API key, an in-memory user map, and auth middleware that returned true for any request carrying an X-Internal header. It was demoed, promoted to production behind the existing gateway, and six months later processes live transactions with all three shortcuts intact and untracked. Which failure mode does this describe?
MCQ-23A fraud scoring PoC is being moved to production. The trained model holds 0.93 AUC on its 40,000-transaction test set and scores in 6 ms at p99. Its strongest feature, merchant_chargeback_rate_30d, is computed by the feature extraction code from an aggregate table that is rebuilt once a day at 02:00 UTC, so the values it yields trail live activity by about 31 hours. Replaying the model over 180,000 live transactions gives 0.58 AUC; recomputing that one feature from same-day rows costs 3 ms at p99 and lifts it to 0.91. The serving layer loads the 380 MB model from disk on every request, 210 ms p99 at 40 rps, and carries a hardcoded database password. Production needs 600 rps at p99 <= 120 ms, secrets from Vault, and per-tenant authorization. Which reuse plan follows from the gap review?
MCQ-11A team trains a VAE on 40M real examples to generate synthetic training data for a rare-class classifier. Real examples: mean 12 events, std 40; 1% exceed 200 events, and those long examples carry 30% of the positive class. The VAE minimizes per-example reconstruction loss. The 40M generated examples have mean 12, std 4, and none exceeds 58 events; the synthetic positive rate matches real. A classifier trained only on the synthetic set scores 0.94 AUC on held-out synthetic data. On real data its recall on positive examples is 0.19 overall, 0.71 on positive examples under 200 events, the same as the real-trained baseline, and 0.02 on positive examples above 200 events. The long examples' positive labels were verified against downstream purchases, and the serving path passes the full example to the model. Which cause fits these numbers?
MCQ-4A team is deciding whether to adopt a read-through cache in front of its primary datastore. The PoC charter, signed before the first load test, set two exit criteria: p99 read latency at 5,000 reads/s under 100 ms, and added infrastructure cost under 10% of the current $48,000 monthly bill. The load test finished at p99 = 95 ms (180 ms without the cache), with the cache adding $5,760/month. The team lead recommends adoption, citing the 47% latency cut. What should the review decide?
MCQ-5A team is choosing a vector index for a 400-million-vector search service. The bar set before the PoC: p99 <= 25 ms at 2,000 QPS with the index resident on 32 GB nodes, and $3,000 a month of infrastructure (a 32 GB node costs $2,400). Run on the team's own nodes and harness, the uncompressed 100 GB index spread over four nodes held p99 = 19 ms at 2,000 QPS with recall@10 = 0.94; the product-quantized index fits in 26 GB on a single node but, fully warmed and resident, holds p99 = 71 ms at 2,000 QPS with recall@10 = 0.62. Which conclusion does this PoC support?
MCQ-12A team is running a 6-week PoC of a feature store. Success requires p99 feature fetch < 20 ms at 5k QPS and cost < $0.50 per 1M fetches. After 4 weeks: p99 is 45 ms, cost is $1.20 per 1M, and the last two weekly experiments each cut p99 by 3 ms with no cost change. $4k of the $20k budget remains. Which stop rule should govern?
MCQ-25A payments team ran a 14-day canary of a new in-process cache on 10% of production traffic. The canary cut p99 latency from 180 ms to 95 ms, but raised the error rate from 0.02% to 0.40%. The quarter's error budget is 0.10% of requests, and the p99 latency SLO is 150 ms. The cache costs $3,200/month. The team's decision review must produce one outcome. Which outcome does the evidence support?
MCQ-20A team evaluates a proposed index change against 12 acceptance criteria. Seven criteria pass. Three fail. Two cannot be measured because the production telemetry needed for them was only enabled after the test window closed. The draft report says the change 'meets all criteria that could be assessed.' How should the two unmeasured criteria be classified in the final report?
MCQ-21A platform team is deciding whether to migrate its primary datastore from a managed service to a self-hosted cluster. A 3-week benchmark, which cost $60k, came back inconclusive. A full 6-month replay would cost $500k in engineer time and delay the migration by two quarters. The team's model: if the workload fits the self-hosted cluster, the migration saves $800k over the 2-year horizon; if it does not fit, it costs $1.2M in rework and incidents over the same horizon. The benchmark leaves the team at 50/50 on whether the workload fits. What should the team do?
MCQ-6A team has run two 6-week pilots of a ranking model. Each ended the same way: offline AUC rose from 0.71 to 0.74 on a 50k-row holdout (p<0.01), the deck called it promising, and the team asked for a third pilot. Nothing has shipped. The third pilot starts Monday with the same charter, the same heuristic baseline, and an executive sponsor already named. Which entry criterion is missing?
MCQ-16A team has not yet decided whether to adopt a new ranking model. They run it for 6 weeks on all traffic from the Irish market (3% of global traffic), with pre-registered success criteria (conversion +2%, error rate <0.1%) and a scheduled go/no-go review. How should this be classified?