FreeExam

Platform system design

A senior platform engineer interview on designing shared build, deploy, compute, traffic, config, telemetry, and flag systems that hundreds of teams rely on, including how they fail and how to migrate onto them.

These prompts are original practice. They are not questions from a named employer or a live hiring loop.

  1. 1.You are asked to design a shared build and continuous integration (CI) system for a company with 900 engineers. What should you ask first?

  2. 2.A director asks you to design a deploy platform for 300 services. Teams deploy anywhere from twice a year to 40 times a day. Why does that spread matter to the design?

  3. 3.You are designing a multi-tenant compute platform. Which question most changes whether tenants can share a cluster at all?

  4. 4.For a new metrics and logs pipeline at a grocery delivery company, which requirement should you pin down before choosing storage?

  5. 5.You are designing a feature flag service. A product manager says it must be highly available. What is the most useful follow-up?

  6. 6.Leadership wants one internal gateway for all service traffic at a ride dispatch company. Which detail should you learn first?

  7. 7.You are scoping a secrets service. Which non-functional requirement matters most for its design?

  8. 8.A monorepo build at a code review tool company has 5,000 actions, each taking about 4 CPU-seconds cold. With a 90 percent remote cache hit rate, about how much CPU time does a typical build need?

  9. 9.At peak, a continuous integration (CI) system starts 400 pipelines an hour, and each uses 20 runner-minutes. About how many runners are busy on average during that hour?

  10. 10.You sized continuous integration (CI) runners for the average load of 133 busy runners. Developers still wait 15 minutes for a runner every weekday morning. What is the likely mistake?

  11. 11.A test suite takes 120 minutes serially. You split it across 20 shards by file count, but wall time is 35 minutes instead of about 6. What should you fix first?

  12. 12.Each test shard spends 4 minutes on setup and 6 minutes on tests. A team asks for 3 times as many shards to go faster. About what wall time should they expect?

  13. 13.A shared remote build cache at a photo backup company returns wrong outputs for a few builds. What design rule most directly prevents this?

  14. 14.Who should be allowed to write to a shared remote build cache that all continuous integration (CI) jobs and developer laptops read from?

  15. 15.The continuous integration (CI) runners at a sensor dashboard company are long-lived machines reused across jobs from all teams. Which risk is most serious?

  16. 16.You move continuous integration (CI) to ephemeral runners that start clean for each job. Build times rise because local caches are gone. What is the best fix?

  17. 17.A monorepo at a payroll export company runs every test on every change. Which approach cuts test time most while staying safe?

  18. 18.Your continuous integration (CI) system runs 60,000 jobs a day. A flaky test fails 2 percent of the time and runs in 3,000 of those jobs. About how many false failures does it cause daily?

  19. 19.A merge queue for a busy monorepo tests each pull request one at a time and lets only 6 merges through an hour. What design change raises throughput?

  20. 20.You are designing a deploy system with automatic rollback. What should drive the rollback decision?

  21. 21.A service handles 2,000 requests per second with a 0.1 percent error rate. The canary gets 1 percent of traffic for 5 minutes. Why is this a weak test?

  22. 22.Which rollout plan best balances speed and safety for a busy stateless service?

  23. 23.A deploy at a news feed company passed all canary checks. An hour after reaching 100 percent, a memory leak crashes every pod. What would have caught it?

  24. 24.A new release adds a database column and code that reads it. Automatic rollback reverts code but cannot undo migrations. How should the change be shipped?

  25. 25.Your deploy system serves 600 services that each deploy about 3 times a workday. Each progressive rollout takes 30 minutes, mostly within an 8 hour day. About how many rollouts run at once on average?

  26. 26.Two teams at a ticket sale company deploy to the same service at once, and one rollout overwrites the other. What should the deploy system provide?

  27. 27.Automatic rollback at a clinic booking company triggered during a regional network blip that hit both old and new pods. What does this show about its signal?

  28. 28.A team wants the deploy system to skip canaries for an urgent fix during an outage. What is a sound design?

  29. 29.Your deploy system rolls out region by region. Which order is usually safest?

  30. 30.A rollback at an airline booking site took 20 minutes because it rebuilt the old version from source. What should change?

  31. 31.A shared cluster has 200 nodes with 32 CPU cores each. Team quotas add up to 9,000 cores. What does this mean?

  32. 32.On a shared cluster, one tenant's job floods the Kubernetes API server with list calls, and other tenants' deploys time out. What should the platform add?

  33. 33.A sensitive tenant needs stronger isolation than namespaces give, but a separate cluster is too costly. What is a reasonable middle step?

  34. 34.On a shared cluster at a music streaming company, one tenant's pods fill node disks with logs and evict other tenants' pods. What should the platform enforce?

  35. 35.A tenant runs latency-sensitive APIs next to another tenant's batch jobs. Both have correct requests. The APIs still slow down. What is the likely cause?

  36. 36.During a capacity shortage, the platform must ensure customer-facing pods get scheduled before batch jobs. Which mechanism fits?

  37. 37.You give each tenant a quota on CPU requests but not on pod count. What risk remains?

  38. 38.Your platform runs 40 clusters with about 1,500 tenant namespaces. How should tenants get the same quotas, roles, and policies everywhere?

  39. 39.Cluster nodes are 85 percent requested but only 25 percent used. Requests come from teams' own estimates. What should the platform build?

  40. 40.You are choosing between one large shared cluster and many smaller clusters. What is the main tradeoff?

  41. 41.A tenant's pods at a fitness app keep getting evicted when nodes run low on memory, though they stay under their limits. What is the likely reason?

  42. 42.An autoscaling cluster takes 4 minutes to add a node. Tenants see Pending pods during sudden spikes. What design change helps most?

  43. 43.You plan a sidecar service mesh for 20,000 pods. Each sidecar uses about 50 MiB of memory. About how much memory does the mesh add?

  44. 44.Each sidecar hop adds about 0.5 ms. A request at a hotel booking site crosses 6 service-to-service calls, each with a client and a server sidecar. About how much latency does the mesh add?

  45. 45.Why would a platform team add a service mesh rather than ask each team to add retries, timeouts, and encryption in its own code?

  46. 46.The mesh retries failed calls up to 3 times at every hop. A call chain has 4 hops. A backend slows down. What can happen?

  47. 47.An edge gateway at a recipe app handles authentication, rate limits, and routing for 80 services. What is the biggest risk in this design?

  48. 48.You want mutual TLS (mTLS) between all services on the mesh. Some legacy services cannot use it yet. How should you roll it out?

  49. 49.Rate limits at a shared gateway are set per instance. You run 10 gateway instances and want a global limit of 1,000 requests per second for one tenant. What is a simple, approximate design?

  50. 50.The mesh control plane pushes every route to every sidecar. With 3,000 services, sidecar memory and push time grow sharply. What is the usual fix?

  51. 51.20,000 pods each read 5 secrets at startup. After a cluster-wide restart, all pods start within 2 minutes. About what read rate must the secrets service handle?

  52. 52.Services at a chat app fetch a database password from the secrets service on every request. What is wrong with this design?

  53. 53.You are designing short-lived database credentials issued by a secrets service. Credentials last 1 hour. When should clients renew them?

  54. 54.A config service pushes a change to 50,000 connected clients at once, and all of them refetch the full config. The service falls over. What should change?

  55. 55.How should a secrets service decide which workload can read which secret?

  56. 56.The secrets service is down for 10 minutes. What should happen to already running services?

  57. 57.A team asks to store a 5 MB machine learning model in the config service. What is the best answer?

  58. 58.A latency histogram has 12 series per label set. It is labeled by 50 endpoints and 5 status codes, across 300 pods. About how many active series does it create?

  59. 59.A team at a game backend adds a user ID label to a request counter. The service has 2 million active users. What happens?

  60. 60.A metrics store holds 1 million active series scraped every 15 seconds. About how many samples per second does it ingest?

  61. 61.Compressed metrics take about 1.5 bytes per sample. You ingest about 67,000 samples a second. About how much storage is that per day?

  62. 62.You keep metrics at 15 second resolution for 30 days, then downsample to 5 minutes. How much does downsampling shrink older data by sample count?

  63. 63.10,000 pods each emit 200 log lines a second at about 300 bytes a line. About how much raw log data is that per day?

  64. 64.Debug-level logs make up 70 percent of 2 TB of daily logs at an insurance claims company. They are rarely searched. What is the best cost move?

  65. 65.Your log pipeline indexes every field of every log line, and indexing is most of the cost. What design change cuts cost most while keeping search usable?

  66. 66.A logging agent on each node sends logs to a central pipeline. The pipeline slows down during an incident. What should the agent do?

  67. 67.Logs are kept hot for 7 days and in cheap object storage for 1 year. Hot storage costs 20 times more per GB. Ingest is 1 TB a day. Which tier holds more data?

  68. 68.How should the platform show teams what their telemetry costs?

  69. 69.A metrics pipeline enforces a per-service series limit. A service hits it during a deploy. What is the best behavior?

  70. 70.A service handles 200,000 requests per second and checks 3 flags per request. Why should flags be evaluated in the client library rather than by calling the flag service?

  71. 71.3,000 service instances poll the flag service every 30 seconds for updates. About what request rate does polling create?

  72. 72.A percentage rollout at a real estate listings site shows a feature to 10 percent of users. Users report it flickers on and off between page loads. What is the likely cause?

  73. 73.When the flag service is unreachable at startup and there is no cached copy, what should a client do?

  74. 74.After two years, a company's flag service holds 9,000 flags, and nobody knows which are still used. What should the platform add?

  75. 75.A flag change turned on a heavy feature for all users and overloaded a database. What should the flag service support?

  76. 76.Which data should a client library send back to the flag service?

  77. 77.Teams at a freight tracking company want to create their own databases through the platform. How should you keep this safe without tickets?

  78. 78.Where should platform policies, such as no public storage buckets, be enforced?

  79. 79.A self-service portal lets teams create services in two minutes. A year later, the platform team cannot change defaults in existing services. Why?

  80. 80.In a platform design, what is the difference between the control plane and the data plane?

  81. 81.The mesh control plane at a weather service goes down for 30 minutes. What should happen to service-to-service traffic?

  82. 82.Your gateway reads routing rules from a central database on every request. Why is this a weak design?

  83. 83.Kubernetes keeps running pods when the API server is down. Which statement describes what still works and what does not?

  84. 84.Which availability target makes sense for a platform's control plane versus its data plane?

  85. 85.A control plane pushes config to 10,000 proxies. A bug produces empty config. What protection should the data plane have?

  86. 86.A Kubernetes controller written by a team lists all 100,000 pods every 10 seconds. API server load spikes. What is the right fix?

  87. 87.You are designing a regional deploy system. Should the control plane be global or one per region?

  88. 88.The internal container registry at a video platform goes down. New pods cannot pull images, and autoscaling fails. What design reduces this risk?

  89. 89.DNS inside the cluster fails, and every service starts failing calls. Which design choice would have limited the damage?

  90. 90.A platform outage at a payroll export company took down 200 services, and teams learned about it from customers. What should the platform team build?

  91. 91.Your platform depends on a cloud provider's identity service. It fails, and engineers cannot log in to fix anything. What should exist beforehand?

  92. 92.After a cluster outage, 5,000 pods restart at once and all connect to a shared Postgres database. It runs out of connections. What should the platform provide?

  93. 93.How should a platform team test that its recovery plans actually work?

  94. 94.The platform's shared message queue at a ride dispatch company backs up during an incident. Producers keep sending at full rate. What design helps?

  95. 95.You design the platform's disaster recovery. Which question decides whether a second region must be active or can be cold?

  96. 96.You must move 400 services to a new platform. With today's process, the team can migrate 10 a week. Automation could raise that to 25. About how long does each path take?

  97. 97.Which services should go first in a migration of hundreds of services to a new platform?

  98. 98.During a migration, how should traffic move from a service's old deployment to its new one?

  99. 99.During a migration of 500 services, some teams keep adding new services on the old platform. What should the platform team do?

  100. 100.A team migrating to the new platform finds its service behaves differently there. How should the platform reduce such surprises across hundreds of services?

  101. 101.You plan to migrate stateful services, such as databases, to a new platform along with stateless ones. What is the main design concern?

  102. 102.The deploy system's service level objective (SLO) is 99.5 percent successful deploys. It runs 2,000 deploys a week. How many failures per week fit the target?

  103. 103.Which service level objective (SLO) best reflects what teams need from a continuous integration (CI) system?

  104. 104.A secrets service has a 99.9 percent monthly availability target. About how much downtime does that allow in a 30 day month?

  105. 105.The platform's metrics system is down during an incident, so teams cannot see their dashboards. What design helps?

  106. 106.How should a platform team decide when to slow feature work and focus on reliability?

  107. 107.You are choosing between a central build service that all teams use and a library each team runs in its own continuous integration (CI). What are you trading?

  108. 108.A colleague proposes building a custom scheduler instead of using Kubernetes for a new compute platform. What is the strongest question to ask?

  109. 109.You can sample traces at 1 percent everywhere or keep all traces with errors or high latency plus 1 percent of the rest. Why choose the second?