Load Testing Senior interview questions and answers

76 questions for a Load Testing (Senior) interview - with correct answers, explanations and links to the docs.

Take this quiz
  1. 1.You build a workload model for a public API. What is the best source of the arrival pattern?
  2. 2.When is a closed workload model the right choice?
  3. 3.How does wrk2 avoid coordinated omission?
  4. 4.A page calls 100 backend services in parallel, each with p99 = 1 s. Roughly what share of page loads see at least one 1 s+ call?
  5. 5.What are hedged requests?
  6. 6.Throughput rises with servers, then falls when more are added. Which model explains this?
  7. 7.How should you size capacity for a service in three availability zones?
  8. 8.You plan to replay production traffic against staging. What is the main risk?
  9. 9.Why mark synthetic load test traffic, e.g. with a header?
  10. 10.How do you make performance thresholds in CI stable enough to block merges?
  11. 11.How do you decide if a latency change between two builds is real?
  12. 12.Why should a request carry a deadline that downstream services respect?
  13. 13.What is a retry budget?
  14. 14.In a queue-based system, how do you know the consumers keep up with the load?
  15. 15.A spike test reaches peak in 30 s, but autoscaling needs 3 minutes. What options do you consider?
  16. 16.A service gets more DB connections in its pool, and latency gets worse. Why?
  17. 17.In Kubernetes, CPU is at 60% of the limit, yet p99 latency spikes. What should you check?
  18. 18.An I/O-bound service autoscales on CPU. What does a load test likely show?
  19. 19.Results look much better than expected. What do you verify first?
  20. 20.Why look at http_req_duration{expected_response:true} and not just http_req_duration?
  21. 21.How do you test that graceful degradation works?
  22. 22.Staging is half the size of production. Can you double staging results to predict prod?
  23. 23.How do you get correct fleet-wide p99 from server-side histograms?
  24. 24.Server histograms have buckets 100, 250, 500, 1000 ms. Why may the reported p99 = 900 ms be imprecise?
  25. 25.Client-side p95 is 900 ms, but app server metrics say 200 ms. Where may the time go?
  26. 26.How should load test results feed into a release decision?
  27. 27.Why combine load with fault injection (e.g. kill a node mid-test)?
  28. 28.What slow resource leaks does a soak test commonly expose?
  29. 29.Before a breakpoint test, what must be agreed?
  30. 30.A multi-tenant SaaS has one huge tenant. What should the load test include?
  31. 31.A CDN purge happens during peak. What risk should a test reproduce?
  32. 32.Why may results depend on the ramp shape, not only on the peak?
  33. 33.How can Little's law check whether a test result is consistent?
  34. 34.Why test critical components in isolation in addition to end-to-end?
  35. 35.A service adds batching: throughput doubles, p50 grows. Is it a regression?
  36. 36.Load testing a serverless API, what limits often appear before your code does?
  37. 37.Which choice of load tool is the most defensible for a team?
  38. 38.How can a workload model represent different user journeys realistically?
  39. 39.Real traffic is burstier than a smooth constant rate. Why does it matter?
  40. 40.A load test of checkout uses replayed requests with the same idempotency key. What happens?
  41. 41.Why sync clocks on distributed load generators?
  42. 42.Under write load, read replicas fall behind. What should the test verify?
  43. 43.Latency jumps for orders of one popular product only. What is a likely cause?
  44. 44.Why watch logging and tracing overhead during a load test?
  45. 45.Why should load test results be reported per time window as well as for the whole run?
  46. 46.How does a load test differ when used for exploration vs as a gate?
  47. 47.Why ensure test data at scale is synthetic or anonymized?
  48. 48.A downstream stub always answers in 5 ms. What is wrong with that?
  49. 49.A rate limiter uses a token bucket with burst 100 and rate 10/s. What should a test check?
  50. 50.Production cache hit ratio is 85%; in the test it is 99%. What follows?
  51. 51.Why should a capacity report state the bottleneck, not just the max RPS?
  52. 52.When is load testing in production justified?
  53. 53.How do you test a long-running database migration's impact?
  54. 54.Why cap the cost of cloud load generators before a big test?
  55. 55.What does it mean that performance is a shared ownership topic?
  56. 56.Why add network latency or loss emulation in some tests?
  57. 57.Why may adding a cache in front of a service create a new risk?
  58. 58.Why model long-lived connections separately in a test (e.g. WebSockets, SSE)?
  59. 59.After a deploy, all WebSocket clients reconnect at once. What should the test check?
  60. 60.Why should load tests also check correctness, not only speed?
  61. 61.Why might you keep a test at peak load for 30-60 minutes rather than 5?
  62. 62.How do you present results to non-technical stakeholders?
  63. 63.What is the danger of a load test that always passes?
  64. 64.Why must think time and pacing be taken from data, not guessed?
  65. 65.You need to emulate 1 million users who each poll every 60 s. What RPS is that?
  66. 66.Which metric best shows that the system sheds load correctly in a stress test?
  67. 67.What is goodput?
  68. 68.Why can queueing inside the load generator corrupt results?
  69. 69.A service is fine at 1,000 RPS from one client, but fails at 1,000 RPS from 10,000 clients. Why?
  70. 70.Why is the end of a load test (ramp-down) worth watching?
  71. 71.Why should a performance budget be part of the definition of done for risky changes?
  72. 72.Why do JVM services often need a longer warm-up in tests than Go services?
  73. 73.The DB is at 40% CPU but has high I/O wait under load. What is the likely bottleneck?
  74. 74.Why add a test that ramps to peak and stays there while a deploy happens?
  75. 75.Why prefer measuring at the edge (user side) for a release gate on latency?
  76. 76.What is the main value of a capacity model built from load tests?