Question bank

4,310 interview questions, answered.

Reference answers, what the interviewer is really testing, how it is graded, and the follow-ups that come next.

Easy 190Medium 2,498Hard 982
Filters · onDifficulty, topic, company
Clear all

982 questions

  1. 961Four aligned ranks report input waits 20,20,140,20 ms and compute 180 ms each. Their collective completes20 ms after the last rank enters. Should networking be the first change?HardAI MLSystem design
  2. 962At10:00 a tenant starts long prompts. At10:02 first-token p95 rises0.8→8 seconds, queue p95 reaches6 seconds, and token-gap p95 stays35 ms. No release occurred. Choose an initial mitigation and recovery checks without claiming a proven root cause.HardAI MLSystem design
  3. 963For user a, changes are(v1,upsert,Ada),(v3,delete,null),(v2,upsert,Amy). User b has(v1,upsert,Bo). Versions are unique and authoritative; arrival order is as listed. Write a current-state query and describe safe incremental behavior.HardDataTechnical
  4. 964A device reports stopped at 09:00, moving at event time09:06 but arriving09:11, and stopped again09:08. The rule requires ten uninterrupted event-time minutes, with the late movement accepted. Give eligibility and delivery identity.HardDataTechnical
  5. 965From the same client, PodIP:8080 succeeds. ServiceIP:80 and the Service DNS name on port 80 both fail. DNS resolves to the expected Service IP. The Service selects the ready Pod but targetPort is 9090. Diagnose the layer and next check.HardSystem designTechnical
  6. 966A team may scale only its approved Deployment from 2 to 8 replicas. A service backend can modify many namespaces. Two requests arrive with the same operation ID but different desired counts. Define authorization and concurrency behavior.HardSystem designTechnical
  7. 967A reachable parser cannot be upgraded for five days. Public uploads and a partner importer both use it; preview of old documents may also parse files. Propose a temporary decision rather than an indefinite waiver.HardTechnical
  8. 968A booking worker crashes after a provider commits but before the local result is stored. The provider's idempotency record expires tomorrow. Write a recovery plan for today and for next week.HardAI MLTechnical
  9. 969Three agents produce a comparison, but one defines an active user as seven days and another as thirty days. Their files have no text conflicts. What should the merge owner do?HardAI MLTechnical
  10. 970An approved draft has two recipients. A worker adds a résumé attachment and a third recipient while the task waits. Specify the resume check.HardSystem designAI ML
  11. 971An agent revision improves 180/200 task passes to 188/200 but creates three unauthorized writes. The old version had none in this set. Give a release decision and the smallest useful investigation.HardDataAI ML
  12. 972You have five visits per patient and want to predict outcomes for new patients at future hospitals. What validation design and limits would you state?HardDataAI ML
  13. 973Only transactions selected for review receive confirmed fraud labels. The team wants to label all unreviewed transactions clean. Explain the problem and a first audit.HardDataAI ML
  14. 974Precision falls after a new region launches, while latency and feature missingness remain stable. Mature labels show the old region unchanged. What is the next diagnosis?HardAI MLTechnical
  15. 975Offline ranking improves, live clicks rise, and completed user tasks fall. How would you decide whether the candidate is useful?HardDataAI ML
  16. 976Loss scale is 512, true gradient norm is 4, and clip threshold is 1. Explain correct processing and the clip-before-unscale error. Use idealized norm arithmetic and ignore numerical epsilon in clipping.HardAI MLTechnical
  17. 977Two ranks have gradient sums 8 and 4 over two and four valid tokens. Compute the global mean and the average of local means.HardAI MLTechnical
  18. 978A reproduction matches sample IDs and logits, but loss and gradient norms are exactly three times larger for a three-example batch. What do you inspect before another full run?HardAI MLTechnical
  19. 979A method improves one benchmark after doubling context and parameter count. What claim is supported and what control tests the proposed memory mechanism?HardDataAI ML
  20. 980You have 600 predictions from six independently trained models. The claim concerns variability across trained models. What is the resampling unit?HardDataAI ML
  21. 981A model is better within easy and hard groups but worse overall. What data do you request and how do you analyze it?HardDataAI ML
  22. 982Your five-seed study supports a method on one language. An interviewer asks you to deploy it in a new language with a different label process. How do you revise the recommendation?HardBehavioralAI ML