Every question on one page: open one, choose an answer, and see why at once. Signed in, your latest answer to each question counts, and you can answer again after 24 hours.
Curated: · Written: · Reviewed:
Reviewed 100Review pending 0
Answers to reviewed questions count toward this role and the overall leaderboard; your latest answer to each one is the one that counts.
MCQ-39A platform team's on-call rotation handled 400 pages last quarter. Classified by where the cause turned out to be: 150 pages were platform faults (median 30 minutes for the platform on-call to prove the platform itself was healthy), 200 were product-code faults (median 25 minutes), and 50 were product-config faults (median 20 minutes). The team wants to quantify how much of its on-call time goes to pages that are not the platform's to fix. How many hours per quarter is that?⌄
MCQ-6An internal developer platform's dashboard shows org-wide deploys per service per week rose from 27.6 in Q1 to 52.8 in Q2. The underlying data: the high-frequency cohort went from 2 services at 90/wk to 6 services at 82/wk; the slower cohort went from 8 services at 12/wk to 4 services at 9/wk. What is the correct reading of the org-wide rise?⌄
MCQ-3A platform team's Q2 review reports adoption of the paved-road pipeline up from 61% to 94% after a migration deadline, and mean time to first deploy for a new service up from 3.0 to 6.0 days. Split by cohort: the 61% who adopted in Q1 average 2.4 days (down from 3.0), the 33% who adopted under the Q2 deadline average 13.2 days, and the 6% who never adopted average 3.0 days. Which reading of the quarter is right?⌄
MCQ-4A platform team is consolidating lead time for changes onto one dashboard for 12 services. Two services report very different figures. Service X: 38 minutes. Service Y: 31 hours. Both run the same pipeline, and merge-to-deploy takes a median of 38 minutes for each. Service X's branches wait a median of 28 hours in review before merge; Service Y's wait a median of 30 hours. Which action makes the two figures comparable?⌄
MCQ-5A platform team proposes to adopt change failure rate, with deployment frequency as its counter-metric, as the platform's headline measure. The proposal reports change failure rate fell from 14% to 3% over two quarters. Over the same period the teams on the platform went from 40 deploys a month to 6, changes per deployment rose from 2 to 14, and mean incident duration went from 41 to 88 minutes. Which statement should the review accept?⌄
MCQ-7A platform org runs a quarterly developer-experience survey scored 0–100. In Q1, delivered through the onboarding checklist, the response rate was 71% and the mean score 74. In Q2 the link moved to a monthly newsletter; the response rate fell to 34% and the raw mean to 66. Reweighting the Q2 responses to match Q1's distribution of team, tenure and role raises the mean to 73. No item wording changed between waves. Which reading of the Q2 result is right?⌄
MCQ-46A platform team reports onboarding time as the median of (repo creation → first production deploy), counting only services that reach production. Last quarter the dashboard read median 2.5 days, p90 9 days. An audit of the 40 onboardings started that quarter found that repos are created the same day the developer is assigned (median gap 0), that the manual access-approval ticket sits a median of 1.8 days, and that 12 of the 40 services never reached production and so never entered the metric. Which change does the audit justify?⌄
MCQ-38An internal developer platform publishes one SLO: 99.95% of 1-minute windows with all platform components reporting healthy. Last quarter it met that at 99.97%. Telemetry for the same quarter shows 4% of `deploy` CLI invocations fail on the first attempt and succeed on retry, median time to a successful deploy is 11 minutes, and 30% of retries are started manually by the developer. The payments team, the platform's largest consumer, promises its customers 99.9% availability and has started running its own scripts against the cluster API to bypass the CLI. Which change to the platform SLO policy best addresses this?⌄
MCQ-1A platform team proposes a self-service deployment pipeline. Nine services each spend about 2 engineer-days a year on deployment work. The pipeline costs 120 engineer-days to build and 6 engineer-days a year to operate. The estate is not expected to grow. Which conclusion should the design review reach?⌄
MCQ-2A platform team is funded as a six-month project: four engineers, the budget ends when the deployment abstraction ships, and they return to product teams. Three product teams want one shared deploy interface; two more are expected to adopt within the year. The proposed design is a config DSL with its own schema, versioning and migration tooling. What should the team build?⌄
MCQ-13A four-person platform team wants to open its provisioning API to every product team. In last quarter's single-team pilot, an engineer who had completed the provisioning training copy-pasted a size value and provisioned 8 instances at 64 vCPU for a staging job; the bill surfaced on the next month's invoice. Which change should the review require before the API opens to everyone?⌄
MCQ-66A platform team of six supports 300 engineers. Last quarter's manual request queue: log access 40/month (worst case a revoked grant, reversible in minutes), quota bumps 22/month (reversible), new-service scaffolding 12/month (a re-run fixes it), production database restores 3/month. In March a restore was run against the wrong instance and destroyed four hours of writes. The team can ship one automation this quarter. What should the team do?⌄
MCQ-67Nine product teams and a 5-engineer platform team own deployment across an org of 62 engineers. Six months in, the platform's deploy path is used by 1 team, the one that co-designed it. Six teams have built wrappers over their old Jenkins jobs and 2 teams still run plain Jenkins. The platform team's last 2 sprints went to 14 one-off requests raised by individual teams, 11 of them from teams not on the path, and next quarter's roadmap is drawn from the platform team's own backlog. Which action fits this evidence?⌄
MCQ-64A platform team logs every intake request and clusters the log quarterly. Last quarter the log held 12 requests from 9 teams, and the team shipped 6 items. A post-hoc review found 4 of the 6 were the same underlying need, a way to run a nightly batch job, shipped as 4 separate implementations. The four log entries read 'need cron on the VM', 'need a scheduler for ETL', 'need a nightly trigger', and 'need a job runner'; each was clustered alone and sized as affecting one team. Which change to the intake process would have prevented this?⌄
MCQ-99A 45-engineer company runs 5 product teams on Kubernetes. Three teams each maintain their own deploy scripts, about 250 lines and diverged; the other two use a vendor's managed CI and container hosting. The duplicated build and deploy work costs the three script-owning teams about 4 engineer-hours each per month. A staff engineer proposes an internal platform team of 3 engineers, 480 engineer-hours a month, to build shared abstractions for all five teams. Which decision does the comparison support?⌄
MCQ-98A platform team of three engineers has spent four quarters on a self-service deployment scaffolding tool. Org-wide time-to-first-deploy moved 2.5 days → 4 hours → 45 minutes → 41 minutes → 40 minutes across Q1–Q4, with the same three engineers on the tool each quarter. Developer complaints now concentrate on access approval, which takes about three days and has no owner. The team is planning Q5. What should the platform lead do?⌄
MCQ-100A platform team agreed, before building a shared config service, to abandon it if fewer than 40% of the 60 backend services had migrated by the end of Q2. At the end of Q2, 21 services had migrated. Adoption then held flat at 21 services through Q4 while the team kept two engineers on the service. In the post-mortem, which statement is defensible?⌄
MCQ-75Six months into a nine-month migration of 240 services from the old platform to the new one, the platform team reviews the numbers. 96 services have moved. 22 services were created on the old platform after the migration started; each must be migrated later and adds about 7 weeks to the tail. On-call pages for the team have risen from 40 to 88 per week, and the old platform's patch cadence has slipped from monthly to quarterly. The team has six engineers. Which action does the data support?⌄
MCQ-74A platform team migrated 100 of 120 services to a new deploy system in 4 months. The last 20 have been in progress for 14 months: 3 of them need a capability the platform lacks, and building that capability changes behaviour for 40 of the 100 services already on the new system, which must then be re-migrated. A second estate of 300 services is next. What should the team do differently?⌄
MCQ-96March: a control-plane database failed over and returned errors for 4 minutes. Deploys stayed broken for 90 minutes across 14 teams, because each team's roughly 20 in-flight pipelines re-queued themselves, whole, every 30 seconds on any control-plane error, and that retry traffic kept the recovering database pinned above 95 percent CPU long after the failover had finished. The platform team's postmortem, in its own repo, covered the failover, the timeline and four platform-owned action items, all closed by May. In June the same 4-minute failover broke deploys for 75 minutes across the same 14 teams. Which change to the March postmortem would have prevented the June repeat?⌄
MCQ-97The shared CI/CD pipeline at a 40-team company runs 1,200 pipelines a week. One stage in it, the platform team's own integration stage, flakes: 20% of runs are marked failed by it, and each of those failures costs the requesting team 30 minutes of an engineer's time to confirm the known flake and re-trigger the run. The platform team has 13 engineer-weeks this quarter and works one workstream at a time. Three candidates are on the table: rebuild the pipeline on a new scheduler, 13 engineer-weeks, with 12 of the 40 teams having filed requests to onboard; fix the flaky stage, root-caused and scoped at 3 engineer-weeks; ship a managed secrets injector, 6 engineer-weeks, with a compliance ticket dated for next quarter. Which item should the team start first?⌄
MCQ-94An internal developer platform team ships a breaking change to the build config schema used by 40 teams' pipelines. They post it in #platform-announcements, the channel that also carries routine notices at about 30 a week, with a three-day migration window. 22 of the 40 teams miss the post and their pipelines fail on cutover morning. Earlier the same quarter, a smaller change emailed directly to the 6 teams it affected, with the deadline and the migration command, was applied by all 6 before the deadline. What should the team change about how it communicates platform changes?⌄
MCQ-44A platform team runs a service catalogue for 500 services. The owner field is typed into a web form when a service is onboarded; a quarterly audit found 230 of the 500 entries named someone who had left or moved teams. The catalogue feeds the incident router, which pages the listed owner: last quarter 40 of 120 Sev-2 incidents were first paged to a wrong owner, adding a median 22 minutes before the right team was engaged. All 500 services are deployed from 12 monorepos, and each service ships a service.yaml whose owner field CI validates against the org directory on every merge. Which change should the team make?⌄
MCQ-45A platform proposal would make documentation part of the interface contract for all 60 services, including the 40 migrated off the old gateway. Last quarter three of four quickstarts shipped examples calling endpoints removed in the same release; all three were for migrated services, and 30% of new-hire onboarding tickets were people following them. Which statement should the review accept?⌄
MCQ-63The platform team owns an internal config library. v2 shipped six months ago with a migration guide, and telemetry shows 37 of 40 services now on v2. Of the three still on v1, two have v2 migrations sitting in code review and one belongs to a service scheduled for decommission in eight weeks. Keeping v1 alive costs the team about two days a quarter in backports and blocks a security fix they need to ship. What should the platform team do next?⌄
MCQ-73The platform team bundled retry policy, per-call timeout, TLS configuration and log shipping behind one routing config object, and all 24 product teams migrated onto it over two quarters. An audit six months later records that 21 of 24 teams override the retry policy and 19 of 24 override the timeout, while TLS configuration is overridden by 0 teams and log shipping by 1; the object now has 22 fields, and median time for a new service to reach production routing is 6 days against 2 days before the library existed. Which change to the interface do these numbers support?⌄
MCQ-65A platform team supports 9 product teams. Two postmortems last quarter each took 4 days to attribute because the bug reproduced only on MySQL 5.7; production and CI run MySQL 8.0. A survey finds 7 distinct docker-compose setups; the 3 teams running MySQL 5.7 locally were involved in both postmortems. The other 6 teams already match CI on MySQL 8.0. Inner-loop test time is p50 75 s and p95 6 min. The team has 2 engineer-weeks this quarter and its goal is to cut repeat attribution delays. Which action should it take first?⌄
MCQ-17A platform team's `ServiceApp` CRD has 38 fields that mirror the Deployment, Service and Ingress objects it generates. Six months in, across 22 onboarded services: the median manifest is 94 lines, teams set a median of 29 of the 38 fields, and 14 services run the same wrong CPU limit, copied from a neighbouring team's manifest. Forty more services onboard next quarter. Which change should the team make?⌄
MCQ-47A platform team ships its service template as a git repo that teams copy into their own repos. The template's liveness probe has changed twice, and the estate now runs three versions:
| Template | timeoutSeconds | failureThreshold | periodSeconds |
| --- | --- | --- | --- |
| v1 | 1 | 3 | 10 |
| v2 | 1 | 6 | 10 |
| v3 | 5 | 6 | 10 |
All three probe the same `/healthz` path. The handler calls a shared config service and takes 4s to answer whenever that service is slow. Last week the config service was slow for 20 minutes. Which services had containers restarted by the kubelet during that window?⌄
MCQ-79A deploy service decodes each service manifest into a Go struct with yaml.v3's non-strict Unmarshal, then fills any zero-valued field with a default: replicas 1, memoryLimit 512Mi. A team submits a manifest containing `replicaCount: 6` and `memory_limit: 2Gi`; the schema's fields are `replicas` and `memoryLimit`. The deploy reports success. At 14:00 traffic triples, the single 512Mi pod is OOM-killed, and p99 rises from 70 ms to 1.4 s. Which pipeline change would have failed this submission before deploy?⌄
MCQ-76Your team ships `deployctl`, a Python CLI, and the onboarding guide tells developers to run `pip install deployctl` into their project virtualenv. Last quarter the platform channel logged 61 tickets about the tool: 44 `ResolutionImpossible` errors from pip, 12 `ImportError`s, 5 `ModuleNotFoundError`s, and 3 about deployctl's own behaviour. It runs cleanly from the platform team's own checkout. Existing Makefiles and CI steps call `deployctl` directly, and about a third of the estate's CI images are minimal builds with no Python or pip installed. Which distribution change fixes the ticket load without touching those call sites?⌄
MCQ-72Release 4.7 of an internal developer platform passed 900 unit tests and 120 component tests, then broke `platform new-service` for every team that tried it: the scaffolding CLI emitted a service-account manifest that the cluster's new RBAC policy rejected, so the deploy step failed. The first report came from a product team two days after rollout. Which change to the release process addresses the cause?⌄
MCQ-43Deployr is your internal control plane: every product team ships through it, and Deployr's own releases are deployed by Deployr. Release v4.2 had passed two-person review and a canary on two tenants, then corrupted Deployr's state store; no team could deploy anything, including the fix. Recovery took 3 hours and needed one of the two engineers who still held raw cluster credentials: the break-glass runbook had last been exercised 14 months earlier. In a drill three months later, nine engineers ran the same runbook in 25 minutes. Which change should the team make next?⌄
MCQ-35Preview environments are created automatically per pull request for 30 product teams. Review cycles run from a day to three weeks, branches are often left in place after merge, teams routinely have more than ten pull requests open during a release, and finance reports cloud spend monthly. A snapshot shows 340 live environments against 41 open pull requests, the oldest still running five months after its branch was deleted. Which design holds up?⌄
MCQ-34Seventy-four teams share one Kafka cluster. The 12,000 USD monthly bill covers the broker fleet, whose size is set by the cluster's busiest hour, Tuesday 19:00-20:00 UTC, when producers push 100 MB/s in total; retention sits inside the disks of that fleet. Finance splits the 12,000 USD evenly today, so each of the 65 teams with no topics pays about 162 USD. Last month:
| Team | Throughput at 19:00 | Monthly bytes | Producer quota | Partitions |
| --- | --- | --- | --- | --- |
| Atlas | 40 MB/s | 30 TB | 40 MB/s | 30 |
| Borealis | 5 MB/s | 60 TB | 80 MB/s | 20 |
| Cobalt | 0 MB/s | 20 TB | 20 MB/s | 60 |
| six other teams | 55 MB/s | 10 TB | 60 MB/s | 90 |
Almost all of Borealis's bytes leave in its 22:00-07:00 batch window. Which allocation should replace the even split?⌄
MCQ-33A payments service writes 400 GB/day to the log platform: 80% per-request access logs, 15% health-check probes, 5% error and warn. Hot retention is 90 days at $0.30/GB-month, about $10,800/month; the archive tier costs $0.02/GB-month and is queryable within minutes. Last quarter, 6 of 7 incident investigations read error and warn lines, and the seventh was a batch failure found 45 days after it happened, from access logs. The team must cut the bill by at least 80% without losing the ability to investigate a repeat of that batch failure. Which plan meets both constraints?⌄
MCQ-32Checkout is asynchronous: POST /checkout writes to the 'orders' Kafka topic and returns 202, leaving a trace of ingress 12 ms plus checkout-api 235 ms, total 247 ms. RUM puts p95 for the completed flow, including the WebSocket confirmation, at 3.9 s. fulfilment-worker exports to the same collector, but its spans arrive as separate traces rooted at the worker, with different trace ids and no parent, starting 3.4 s after the order was produced and running 210 ms. The mesh injects W3C traceparent on every HTTP hop, and the collector head-samples at 1%. On-call wants one trace per checkout that names the hop that spent the 3.6 s, with no manual correlation. Which change produces that?⌄
MCQ-8A platform team's service template is on version 6. Its adoption dashboard counts a service as adopted if it has ever run the scaffolder, and reports 90% across 120 services. An audit finds 41 services on a template version released in the last two quarters, 67 on a version at least three releases old with hand-edited pipeline config, and 12 that never onboarded. Leadership wants adoption at 100% this quarter. Which action addresses the cause of the gap between the dashboard and the audit?⌄
MCQ-11A platform team is deciding how to offer message queues to 40 product teams. Their deploy CLI precedent: last quarter, 12 scheduling incidents; with generic errors, median diagnosis was 74 min and 61% escalated to the 2-person platform team; a pilot that surfaced the scheduler's own message cut that to 19 min and 9%. The queue abstraction's error-mapping layer is estimated at 1.5 engineer-quarters; the team has 1 engineer-quarter of spare capacity this quarter. Which plan fits?⌄
MCQ-12A platform team replaced a 22-step manual deploy with a self-service tool. Across 12 product teams, deploy steps fell to 5, but platform support tickets rose from 40 to 260 a month with the same 4 engineers, median ticket response went from 3 hours to 2.5 days, and 3 teams wrote their own deploy scripts. Before onboarding 8 more services, which conclusion does this data support?⌄
MCQ-14An internal developer platform's portal exposes a Delete service action that runs terraform destroy on the service's stack. The database and its automated backups sit in that same stack, so one delete takes both. Last month a team deleted what it took to be an unused staging service; the stack actually backed a customer-facing preview environment. The team noticed when the preview failed to deploy three days later, and recovery took 11 hours from an offsite copy because the restore path had never been run. The platform carries 300 services across 40 teams and sees about 12 deletions a week, most of them ephemeral preview environments that teams recreate within the hour. Which deletion policy should the platform adopt for the portal's delete action?⌄
MCQ-9A platform team runs a paved-road pipeline with nine mandatory controls, including artifact signing and audit logging. Exit telemetry shows 44 teams on the path, 6 using a component-level escape hatch, and 1 team that rebuilt the pipeline and now retains 3 of the 9 controls. Five of the 6 hatch users exit at the same step: the deploy step rejects their runtime. The team has one quarter and cannot support every runtime. What should they do?⌄
MCQ-10Your platform team ships a golden-path service template. v3 changes the config schema in a breaking way; the announcement gave teams a 30-day window and no codemod. At day 90, 41 of 60 services are still on v2, teams report the manual edit takes about two engineer-days each, and the platform team is maintaining both schemas while the docs say only v3 is supported. What should the platform team do next?⌄
MCQ-58A payments service runs 24 pods behind a rolling update with maxSurge 25% and maxUnavailable 0; a full rollout takes 8 minutes. Release 412 renames column amount_cents to amount_minor and ships the rename as a single pre-deploy migration. During the rollout, writes from the not-yet-replaced pods fail with 'column amount_cents does not exist', and the write error rate averages about 50% for the full 8 minutes. Which change removes the failing window?⌄
MCQ-22A payments API serves 12,000 requests an hour. Its canary policy sends 5% of traffic to the new build for 10 minutes and promotes it if the canary's error rate is not significantly worse than the concurrent control's at 95% confidence. The service's baseline error rate is 1%. Last week a build that doubled the error rate to 2% was promoted and ran for six hours before anyone noticed. Which change to the canary policy would have caught that regression?⌄
MCQ-21A service's deploy job runs `docker build` separately in each environment. Last Tuesday the staging deploy of commit 9f2c1a passed its smoke tests; the production deploy of the same commit crashed at startup with `NoSuchMethodError`. The build logs show staging resolved `lib-common 2.4.1` and production resolved `lib-common 2.4.3`, published 40 minutes before the production build. Which change removes this class of failure?⌄
MCQ-48A monorepo's CI pipeline runs on every push and takes 22 minutes end to end. Across last week's 400 builds the stages averaged: dependency install 14 min, compile 5 min, test 3 min. The dependency cache reported a 0% hit rate on all 400 builds, though the lockfile changed only 6 times. The cache key is `deps-${CI_PIPELINE_ID}-${sha256(lockfile)}`. Which change cuts the most pipeline time?⌄
MCQ-49A CI platform runs 12,000 jobs a day for 240 repositories, and 9% of jobs fail. The team ships a rerun-on-failure button; a month later 71% of failed jobs are rerun and 62% of all reruns come from the same 40 tests. Instrumentation shows 37 of those tests fail only on the 8-core runner class and pass on the 16-core class, while the other 3 fail on both classes when the shared Postgres fixture is under load. A payments regression shipped to production after its failing job passed on the fourth rerun. Which action does this evidence support?⌄
MCQ-50A registry holds 3.1 TB of image tags across 40 services. A cleanup job is proposed that deletes every tag not pulled in the last 90 days; its dry run reports it would remove 61% of tags and free 1.9 TB. The billing service deploys quarterly: its current production image is 96 days old and the release before it is 187 days old. The checkout service deploys about 40 times a day. What should the platform team do?⌄
MCQ-93A platform team proposes a dependency-scanning rollout for 400 services that share 12 base images. The goal is to cut the time to remediate findings exploitable from untrusted input. No team has headcount for dedicated triage, exploitable findings have come from both the shared base images and service-level dependencies, and one in a public-facing service has been open five weeks. Which statement should the design review accept?⌄
MCQ-92Your platform team owns `base:2024.06`, used by 200 services. A CVE in its OpenSSL package is fixed in `base:2024.07`, which you publish and announce. Six weeks later the registry shows all 200 services have pulled `base:2024.07` at least once, but production telemetry shows 116 services still run containers built from `base:2024.06`; 84 have rebuilt. The platform controls the CI/CD pipelines for 180 of the 200 services; the other 20 are owned by teams with their own pipelines. The dashboard says 'CVE remediated' because the patched image exists. The next ten services onboard in two weeks. What should the platform team change first?⌄
MCQ-68A platform team's paved road applies an encryption-at-rest control when it provisions a storage bucket, and a separate policy also requires encryption at rest. An audit of the 210 buckets in the estate found 4 unencrypted. All 4 were created through the cloud console during incidents; all 206 created through the paved road were encrypted. Which change most reduces the number of unencrypted buckets next quarter?⌄
MCQ-70Your internal developer platform exposes 12 capabilities: artifact pulls, container registry, build cache, preview environments and others. Last Tuesday a regional object-storage degradation broke artifact pulls for 40 teams. The platform status page stayed green: its only health check pinged the provider's control-plane endpoint, which was healthy. The provider's own status feed did report the degradation, but 20 minutes late and only as "Object Storage: degraded in us-east-1". The same object storage also backs the container registry and the build cache, and last month a different provider service degraded for 40 minutes with the status page green again. The team has one sprint to change how the platform handles provider incidents. Which change best addresses the failure?⌄
MCQ-71A platform's control plane runs on a single Postgres instance that holds the provisioning queue, the deployment target-version table, and the audit log. Rollback reads the target version from that table before reverting; log read is served from a separate object store. A bad schema migration corrupted the target-version table, and the standby had already replicated the corruption, so every read of the table returned wrong versions. Provisioning stopped (40 teams blocked) and rollback stopped too, so teams could not revert the change that had triggered the incident; log read kept working. Which single change would have kept rollback working through that incident?⌄
MCQ-23A payments service deploys with Kubernetes manifests that reference `payments-api:v2.4.1`. During an incident, the team rolls back to that tag; the rollback runs digest `sha256:9f...`, which contains a cherry-picked commit from `release/2.4` that was not in the original v2.4.1 build. The original digest `sha256:3c...` is no longer tagged. Error rate stays at 7% instead of returning to 0.2%, and the review spends 2 hours comparing logs to the wrong commit. The registry supports immutable tags, but the release process must keep re-pushing the same tag for hotfixes, so immutable tags cannot be enabled. The admission controller verifies image signatures but not the source commit they attest to. Which change best prevents a rollback from running a different commit while keeping the hotfix process?⌄
MCQ-24Forty teams share one production Kubernetes cluster. Last quarter a control-plane upgrade took three attempts across six weeks because every team had to share one maintenance window, and the payments team's admission webhook blocked one attempt for everyone. In the same period the analytics batch job's memory spike pushed API server p99 from 40 ms to 180 ms for about 20 minutes, and readiness probes for six unrelated teams failed. Payments and identity must now evidence a documented security boundary for an audit and want to upgrade on their own cadence. Which change should the platform team make?⌄
MCQ-25A payments service runs 12 replicas across a pool of 6 nodes, each with 8 vCPU. Every pod is set to requests: cpu 200m and limits: cpu 500m. Node CPU utilisation across the pool is 55%. Over the last hour p99 request latency rose from 60 ms to 400 ms. Per-pod metrics show mean CPU 175m, p99 CPU 480m, and container_cpu_cfs_throttled_periods_total / container_cpu_cfs_periods_total at 0.31 on 9 of the 12 pods. Which change is most likely to bring p99 back down?⌄
MCQ-26A Java 21 checkout service runs 6 pods behind an HPA v2 targeting 70% CPU utilization (min 6, max 24). It is IO-bound on a payments API call. During an unannounced 12-minute flash sale the fleet never leaves 6 pods: per-pod CPU peaks at 34%, memory holds flat at 1.1 GiB of a 2 GiB limit, requests in flight per pod climb from 15 to 480, and p99 latency goes from 90 ms to 3.8 s. Which change makes the HPA scale during the next such sale?⌄
MCQ-36A checkout service runs 20 pods at 50 rps each on nodes that hold 4 pods. A campaign will lift traffic from 1,000 rps to 3,000 rps within 30 seconds and hold it for 10 minutes. The HPA targets 70% CPU and a new pod is ready 40 seconds after it schedules, but the cluster autoscaler takes about 4 minutes to provision a node. The team must serve the burst without dropping requests. Which action closes the gap?⌄
MCQ-31A 40-node cluster runs Calico, which enforces NetworkPolicy. The `payments` namespace holds 12 pods. The `frontend` namespace carries the label `team=frontend` and calls those pods; the `monitoring` namespace scrapes `/metrics` from them; the payments pods themselves resolve DNS in `kube-system` and connect to Postgres in `data`. No NetworkPolicy exists yet. An engineer applies:
```yaml
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
name: allow-frontend
namespace: payments
spec:
podSelector: {}
policyTypes: ["Ingress"]
ingress:
- from:
- namespaceSelector:
matchLabels:
team: frontend
```
Which statement describes what the cluster does next?⌄
MCQ-20A platform runs 200 Applications under Argo CD with automated sync and self-heal enabled. One Deployment declares replicas: 4 in Git and is also managed by an HPA that scales it between 4 and 20. During a traffic spike the HPA scales the Deployment to 12; 90 seconds later self-heal reverts it to 4, p99 rises from 180 ms to 2.4 s, and the incident runs 25 minutes longer. Which change fixes the cause without giving up drift correction for the other Applications?⌄
MCQ-16An operator owns a BackupPolicy custom resource. 200 of them exist in the cluster, and the controller watches BackupPolicy objects and enqueues a CR whenever one changes. Every reconcile writes status.lastChecked = <current RFC3339 timestamp> plus one count field, and those status PUTs are the only API-server writes to these CRs. Over a one-hour window in which no BackupPolicy spec changed, the apiserver records 239,000 PUTs to the CRs' /status subresource and the operator logs 240,000 reconcile calls; the controller runs as a single pod with MaxConcurrentReconciles=1, and successive PUT bodies for the same CR differ only in that timestamp. Which change stops the loop?⌄
MCQ-18A `TenantDatabase` operator writes status only after a successful reconcile; on a failed one it returns the error to controller-runtime, which requeues with the default exponential backoff (5 ms doubling to a 1000 s cap). The developer's pipeline gates on `Ready: True`. At 09:00 a developer raises `spec.storageGB` from 100 to 200, which sets `metadata.generation` to 7. At 09:04 `kubectl get tenantdb -o yaml` shows `status.conditions[Ready]=True` with `status.observedGeneration: 6`, and the cloud console still reports 100 GB. What is the correct reading of that status?⌄
MCQ-19A platform team runs a validating admission webhook on its shared cluster. The documented standard requires every Deployment to set CPU and memory requests and limits. The webhook has run in audit mode for a month and logs 3 violations across 140 Deployments, all owned by the payments team. About 60% of Deployment applies go through the team's CI pipeline; the rest come from kubectl and a GitOps controller. The cluster sees about 90 applies per week, and the change policy caps policy-rejected applies at 5% of that weekly volume. The team wants the rule to actually hold. Which action should it take?⌄
MCQ-80Product teams' workloads on the shared Kubernetes cluster started failing admission in the namespaces that already enforce quota, and the pods that did schedule were the first evicted under node pressure, a week after the platform team standardised deployment manifests across the estate. Which platform-level change addresses the cause?⌄
MCQ-40A platform team runs Terraform against an S3 backend. A cost cleanup three weeks ago deleted the DynamoDB lock table, and nobody noticed because CI's concurrency limit had been serializing applies. Last Tuesday two merges landed 40 seconds apart and both CI jobs ran `terraform plan -out=plan.bin` then `terraform apply plan.bin` against the same state. Apply A replaced `aws_instance.web`: the AMI changed, so it destroyed the old instance and created a new one with a new ID, and wrote state at 10:04:20. Apply B's saved plan, computed from the state as it was before A started, showed no change to `aws_instance.web`; B wrote its own state file at 10:04:40. What does the state file now hold for `aws_instance.web`?⌄
MCQ-41A platform team's `service` Terraform module wraps an ECS service and task definition. It exposes 31 inputs: one for nearly every argument of the two resources. It is called from 26 places across the estate. A script that hashes each call's resolved input map reports 24 distinct hashes. Last quarter a security requirement (read-only root filesystem) took 26 pull requests and three weeks; two call sites were later found to have disabled it by passing `container_definitions` wholesale. The team has one engineer for the next quarter and wants the security requirement after that, whatever it turns out to be, to cost a single edit in the module. What should that engineer do?⌄
MCQ-81A platform team replaces its hand-maintained apply order with a schedule derived from declared dependencies. A component starts the moment every component it depends on has finished, any number of components run at once, and each takes the full time shown: vpc takes 4 minutes and has no dependencies; kms takes 5 minutes and has none; rds takes 6 minutes and depends on vpc and kms; redis takes 2 minutes and depends on vpc; api takes 5 minutes and depends on rds and redis; worker takes 4 minutes and depends on redis and kms; batch takes 3 minutes and depends on rds; edge takes 2 minutes and depends on api. With unlimited runners, what is the minimum wall-clock time for the whole apply?⌄
MCQ-82A checkout service broke somewhere in the 1,024 commits merged since the last good release. The team's regression test fails on the current release, but on the last good release it also fails about 1 run in 5. They bisect with a single test run per commit; it reports commit #384 as the first bad commit, and reverting #384 leaves the failure rate unchanged. What should they do next?⌄
MCQ-84A cache fleet shards keys across 10 nodes with shard = hash(key) mod 10 and runs at a 95 percent hit rate. To absorb growth an engineer changes the modulus to 11 and restarts the clients; within a minute the fleet-wide hit rate is 12 percent and origin reads jump to the large majority of traffic. Which account of that collapse is correct, and what should the team do before the next capacity move?⌄
MCQ-89auth-lib is a shared library in a monorepo, and its dependency graph is built from lockfiles. auth-lib is depended on directly by checkout, search and billing. checkout is depended on by web and mobile, and web is depended on by partner-portal. search is depended on by admin. billing, mobile, admin and partner-portal have no dependents. auth-lib itself depends on crypto-utils, and five other services, none of which depend on auth-lib, depend on crypto-utils. A behaviour change lands in auth-lib. Which set of services must be re-tested for blast radius?⌄
MCQ-91A fleet of 60 hosts was built from image v42 nine months ago. Since then engineers applied 14 security patches directly to running hosts; the runbook records 6 of them, and 11 hosts received all 14 while the other 49 received between 3 and 12. A critical CVE in libssl now needs fixing, and the team must be able to rebuild any host to a known state within an hour. Which plan meets that requirement?⌄
MCQ-42A team's Terraform plan reports 'No changes' against an AWS account. A security audit finds 4 security groups with 0.0.0.0/0 inbound rules on port 22, created 14 months ago by an engineer working in the console during an incident. The state file lists 412 resources; the account contains 473. What should the team do?⌄
MCQ-27A platform team runs 74 services on Kubernetes, with secrets in Vault. Today every pod gets the same Vault token (a 12-month TTL, policy `read` on `secret/*`) injected as an environment variable. The team replaces it with per-pod workload identity: each pod presents its projected service-account token to Vault's Kubernetes auth method and receives a Vault token with a 15-minute TTL scoped to that service's own path, re-exchanging when it expires. A pod is compromised at t=0 and the attacker copies its memory and filesystem, which include the Vault token and the projected service-account token. What can the attacker read, and for how long?⌄
MCQ-28A platform team rotates one shared database credential on a 30-day schedule for 12 services, and the cadence is contractual. Last rotation, three services kept using the old value and failed authentication 40 minutes later when their connection pools recycled, after the rotation had been recorded as successful. Pools recycle on their own schedule, some after days. Ten more services onboard next quarter, deploys are continuous, and there is no shared maintenance window. Which approach should the team adopt?⌄
MCQ-29A four-person platform team runs 60 services on one shared long-lived cloud key. The key leaked from a public repository; the audit log showed 400,000 calls that could not be attributed to any of the 60 workloads, and the response was to rotate the key and restart all 60 services. The cloud provider supports OIDC workload federation, and the team has no capacity to run and page on a credential-issuing service of its own. Which statement should the review accept?⌄
MCQ-30An internal developer platform issues one shared `deploy` role to 30 services. A 90-day access-log review shows 40 permissions granted, 18 exercised, and 22 never exercised. Among the 22 are `s3:DeleteObject` on the production artifacts bucket, `rds:CreateDBSnapshot` used by a quarterly disaster-recovery drill that last ran 95 days ago, and `s3:GetObject` on the compliance-export bucket used by a quarterly export that last ran 100 days ago. What should the platform team do?⌄
MCQ-62Your platform team supports 180 services and needs a secrets-management capability. Building a broker on the cloud KMS: with HA, key rotation, audit logging and the SOC 2 evidence your auditors require: is estimated at 9 engineer-months, plus about 2 engineer-months a year to maintain. A vendor's managed secrets product costs $40k/year, stores secrets in a documented format with a full export API, and ships its own client libraries. Secrets management is not something your company competes on. Which choice fits this situation?⌄
MCQ-69A 40-service platform logs every production change under the deployer's service account. After a deletion, investigators joined three systems by timestamp and took two days to name the person. The platform handles 12,000 changes/day, and security requires attribution to the initiating human for 400 days. The team proposes one change. Which proposal meets the requirement?⌄
MCQ-95A platform team owns deployment tooling for 60 services. Last quarter, 9 of the 12 break-glass production database grants came from one product team, each lasting the full 2-hour maximum, and the audit log shows the same three diagnostic queries run every time. The team now asks for standing read access for its 40 on-call engineers. Security policy requires production data access to be audited and time-bounded. What should the platform engineer do?⌄
MCQ-51A platform team runs POST /v1/instances, which provisions a database instance and returns a server-assigned ID. Clients time out at 30 s and retry; a nightly reconciliation job also replays any request whose response it never recorded, so a replay can arrive up to 24 hours after the first attempt, from any of three regions. The SDK mints a fresh UUIDv4 request ID on every attempt and stamps the body with the current time, so no two attempts at one logical request share a body. Customers routinely run several instances of the same size in one account. Which design makes the retries safe?⌄
MCQ-52A provisioning endpoint runs synchronously: the handler creates the resource and returns 201. Measured p99 is 240 s, p50 is 8 s. The request path has a 60 s idle timeout on the shared load balancer (fixed by the edge team, cannot be raised) and a 120 s client timeout. Last week 4,000 requests hit the endpoint; the server logged 4,000 completed provisions, the load balancer logged 3,600 504s, and support found 900 duplicate resources. The client SDK is embedded in third-party integrations and cannot be changed. Which change fixes the duplicates without breaking the integrations?⌄
MCQ-53An inventory service exposes `GET /v1/resources?offset=&limit=`, sorted by `updated_at` descending, with `limit` capped at 1000 server-side. The collection holds 2.4M resources. A nightly reconciler walks it page by page and marks any resource it does not see as deleted. Last night's run took 3h12m; during it 1,400 resources were created and 1,500 were updated or deleted, and the reconciler processed 40 resources twice and never saw 31 others, tearing down their alerts. The walk must not hold a database transaction open for the length of the run. Which change makes the reconciler see every resource that existed for the whole run exactly once?⌄
MCQ-54A shared provisioning control plane in an internal developer platform degrades above 2,000 requests/hour: p99 climbs from 120 ms to 9 s and provisioning calls start failing. Forty internal services call it, each authenticating with its own service account, and all CI egress leaves through three shared NAT addresses. A month of telemetry shows the busiest legitimate caller peaking at 22 requests/hour. Retry logic comes from a helper vendored into 40 repositories: it retries immediately on 5xx, with no backoff and no attempt cap, and surfaces 4xx to its caller. Last Tuesday one pipeline's helper entered a retry loop and drove 41,000 requests/hour for 90 minutes, during which provisioning failed for all 40 teams. Which policy should the platform team ship?⌄
MCQ-55A platform team delivers webhooks to 40 consumer teams. Each event is HMAC-signed and delivered at-least-once with 3 retries spanning 90 seconds, after which it is dropped; payloads carry nothing a consumer could use to order or dedupe them. Last month a consumer's endpoint was down for 40 minutes during a rolling deploy. 61 events were dropped, and the team only noticed two weeks later when their inventory count disagreed with the platform's. They have asked the platform to replay the 61 events. Which change should the platform make?⌄
MCQ-56A platform control plane authorises deploy requests through two decision caches. The API gateway caches each (principal, resource, action) allow/deny for 300 s. The policy service behind it caches the same key for 60 s. The policy service reads role bindings from a replica that trails the primary by up to 30 s. Nothing invalidates a cached entry when a role binding is revoked; entries expire only when their TTL runs out, measured from when the entry was written and not refreshed by hits. Security requires that a revoked grant stops working no later than 240 s after the revocation is committed to the primary. Which is the largest API gateway cache TTL that meets that requirement if every other setting stays as it is?⌄
MCQ-57A provisioning service enqueues jobs that 30 workers drain by calling a cloud API limited to 10 requests/sec per account; the queue is bounded at 20,000 jobs. During a 3-hour API outage the queue filled to 14,000 jobs. When the API recovered, the workers drained at full speed (about 300 requests/sec), the API returned 429s, and the account was throttled for a further 2 hours. Which change fixes the recovery?⌄
MCQ-78A platform client library calls a downstream service that serves about 200 requests/s before it saturates. Each of 400 client instances retries a failed call every 1 s with no cap on attempts. A 2-second downstream blip made all 400 clients retry in lockstep, and the retry load kept the downstream saturated for 6 minutes after the blip had cleared. The platform's SLO requires that a single transient failure not surface to the caller. Which change to the retry policy fixes the sustained saturation?⌄
MCQ-83A gateway enforces a per-key limit of 1,000 requests per minute. A nightly batch client must be able to send its 1,000 requests in a single burst, and the downstream database saturates if more than 1,000 requests arrive in any 10-second window. The team must keep the sustained limit at 1,000 requests per minute. Which rate limiter meets all three requirements?⌄
MCQ-87A provisioning service calls CreateBucket on a cloud control-plane API, then immediately calls GetBucket to confirm; when GetBucket returns 404 it calls CreateBucket again. Under light load it never duplicates. Under heavy load, read-after-write lag is p50 210 ms and p99 2.4 s, and roughly 1 in 30 runs creates two buckets. Which change removes the duplicates?⌄
MCQ-88A request traverses ingress → control plane → provisioner → cloud API. Each hop sets its own fixed 30 s timeout and retries once on timeout. The user-facing budget is 30 s, and the provisioner's work is variable: it can legitimately take 20 s. Under an intermittent cloud-API slowdown, ingress returns 504 at 30 s while the provisioner and cloud call keep running to completion, so one slow call holds resources at all four levels. Load is otherwise normal. Which change addresses the cause?⌄
MCQ-85A platform worker consumes jobs from an in-process queue that HTTP request handlers fill on the way out of a request; the handlers must not block, and each job is best-effort and re-derivable from the request that produced it. In a load test where a downstream dependency slowed from 5 ms to 200 ms per call for 30 minutes, the unbounded queue grew to 2.1 GB; the worker was OOM-killed and 40,000 queued jobs were lost with no record of what they were. Which change should the design review require?⌄
MCQ-86An internal developer platform runs its controllers as 3 replicas for availability, using client-go's leaderelection with LeaseDurationSeconds=15, RenewDeadline=10s and RetryPeriod=2s. One reconcile of a large tenant takes up to 40s and creates cloud resources with generated names. During a 90-second network partition from the API server, the old leader's lease expires and a new leader is elected; when the partition heals, several resources exist twice. Which change removes the duplicate writes?⌄
MCQ-90Users open a platform dashboard cold on a 10 Mbps link and need the globally sorted first page quickly. The dataset is 200,000 rows, 140 MB as JSON, and the current build sorts it in the browser, taking 45 s to first paint and holding 2 GB in the tab. What should the team do?⌄
MCQ-59A platform team is about to raise the autoscaling ceiling for 12 services that share one Postgres instance with max_connections = 500. Each service's connection pool defaults to 20, and the new ceiling is 30 replicas per service. Which plan holds up?⌄
MCQ-60The platform team is standardising feature flags before onboarding ten more services. Flags cover per-service release toggles and a handful of estate-wide kill switches: one disables a payment path in every service that calls it. The evaluation service has been unreachable for 40 minutes twice this quarter. Which client-side contract should the platform mandate for every service it onboards?⌄
MCQ-37Your team runs the shared feature-flag service. At 14:02 a config push makes flag evaluation return stale values for roughly 30% of evaluations; reads succeed, so nothing errors. The rollback pipeline takes 25 minutes and cannot be shortened. Sixty services evaluate flags on every request. In the last incident of this shape, the first consumer rolled back its own deploy 8 minutes after the first stale read, and 14 teams had opened their own investigations before a notice went out 35 minutes in. It is 14:05; you have a hypothesis but no confirmed root cause. What do you do?⌄
MCQ-77A platform job reconciles 4,000 resources against a cloud API, one GET and one PATCH each, so 8,000 calls. The account's published limit is 50 requests/s with a burst of 100, and two other teams run provisioning jobs against the same account overnight. Calls returned in about 450 ms at every concurrency tested.
| Configuration | Wall clock | 429 responses | Other teams' jobs |
| --- | --- | --- | --- |
| 1 worker | 60 min | 0 | unaffected |
| 20-worker pool | 3 min | 0 | unaffected |
| 4,000 concurrent tasks | 22 min | 2,130 | blocked 25 min |
The 20-worker pool was replaced by the 4,000-task fan-out, on the argument that 22 minutes still beat the serial 60. The other teams' jobs must keep running. Which change should the team ship?⌄
MCQ-61A platform team asks the review board to approve automatic cross-region failover for the internal deployment platform. The control-plane database replicates asynchronously to the secondary with a measured 40-minute lag at peak; the artifact store is replicated nightly; secrets are synced continuously. The runbook exists but has never been executed. Which decision should the board make?⌄
MCQ-15An internal payments API is on v1. Per-caller instrumentation was switched on the day the 16-week sunset window was announced; it counts distinct callers per version, split by whether the calling service has a registered owner in the service catalog. Observed callers still on v1:
| Week | Registered owner | No owner |
| --- | --- | --- |
| 0 | 300 | 110 |
| 4 | 168 | 108 |
| 8 | 54 | 104 |
| 12 | 21 | 99 |
| 16 | 9 | 96 |
The window expires at week 16. What should the team do about removing v1?⌄