Skip to content
Tech Interview Prep
Technical interview guide

Golden Paths & Paved Roads

The officially supported, easy default way to build and deploy: that's still possible to deviate from when genuinely needed.

Read
38 min
Practice MCQs
25
Interview Q&A
25
Edition
v3
Editorial status
Reviewed
Relevant for
Platform Engineer

Scope: AWS Internal Developer Platform Prescriptive Guidance, Microsoft Platform Engineering guidance, Google Cloud platform engineering guidance, and Backstage stable documentation current 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-18A platform team's quarterly dashboard reads "Golden-path adoption: 53%". The team's goal is to move production services off the legacy Jenkins pipeline onto the golden path. The numbers behind the dashboard: the org has 400 engineers, of whom 300 own services and hold deploy rights; the other 100 work in data science, design and program management and never deploy. Of the 300 service owners, 150 have their primary production service deploying through the golden path, 60 have run the golden path only for a side project or a non-production service while their primary service still deploys through legacy Jenkins, and 90 have not used it at all. Which figure should the dashboard report as golden-path adoption?
MCQ-19A platform team's Q3 review reports 1,240 executions of the internal service template, up 38% on Q2, and calls the golden path a success. The telemetry behind it: 610 executions came from 180 existing repositories re-running CI after a template version bump; the other 630 each created a distinct repository, of which 214 have had no commit since creation and 96 have commits but have never run a deploy pipeline. The fastest-growing tag in the support queue is "how do I stop using the template". Which reading of the 1,240 is correct?
MCQ-17A platform team ships a golden-path template for new services and lets teams adopt it or keep their existing setup. Six months in, the dashboard shows 340 services created from the template with a median 2.1 days from repo creation to first production deploy, against 8.4 days for the 610 services built by hand in the same period. Leadership reads this as a 75% lead-time cut and asks the team to justify the template's cost. What should the team do before accepting that number?
MCQ-2The platform team can pave exactly one golden path this quarter. They pulled 90 days of telemetry for the five journeys developers run most: | Journey | Runs/week | Median time per run | Teams affected | |---|---|---|---| | Bring up a local dev environment | 200 | 45 min | 12 | | Deploy to staging | 500 | 12 min | 12 | | Provision a production database | 2 | 4 days | 6 | | Create a new service repo | 15 | 6 h | 10 | | Rotate a service credential | 40 | 2 h | 14 | The figures are illustrative. Which journey should they pave first?
MCQ-1A platform team owns one default CI/CD pipeline for stateless HTTP services. Last quarter, 32 of the 40 HTTP service teams used the default; each deployed 10 times. The default logged 0.30 tickets per deploy, and 72 of its tickets came from one manual migration approval that fires on every deploy. The other 8 HTTP teams run bespoke pipelines: 4 deploys each at 2.0 tickets per deploy. Separately, 14 teams run Spark batch jobs with no supported default: 2 jobs each at 3.0 tickets per job. Only 5 of those Spark teams repeat the same two-job pattern; the rest are diverse. The team can land one change this quarter. Which change removes the most quarterly platform tickets?
MCQ-14A platform team's telemetry shows 40% of deploys use the golden-path pipeline. Of the 60% that don't, 22 services in the payments org run a near-identical Kafka-to-Snowflake sink each team built itself, 14 ML services bypass the standard container build because the golden path rejects images over 2 GB and their model images are 6–8 GB, and 3 teams run pipelines sharing no pattern with anyone else. What should the platform team conclude?
MCQ-25A platform team maintains a paved-road template for Go services. Six weeks after releasing template v3, which adds automatic canary analysis and a generated deployment manifest, the team measures: 200 services total; 120 still on v2; 80 on v3. Of the 80 on v3, 52 have hand-edited the generated manifest, and 19 of those 52 have rolled back after a later template patch because their manifest was stale. Separately, 14 teams created services outside the template, and 11 of those services have no owner label. Which conclusion about the paved road is best supported?
MCQ-3A platform team runs a paved road for service deploys: 25 minutes from commit to a staging deploy, with canary, audit logging and one-command rollback. The raw path, hand-edited Helm plus kubectl, takes about 3 hours and has neither canary nor an audit trail. Six months in, 34 of the team's 120 services deploy through the paved road. Of the 86 that do not, 61 say the road cannot do something they need (a custom autoscaling policy, a Postgres extension, a second region), 12 did not know it existed, and 13 prefer their own tooling. Four of last quarter's five Sev-2 incidents came from raw-path services. Two engineers have one quarter to lift adoption, services whose production deploys run through the paved road, above 75%. Which action gets closest?
MCQ-23A platform team's internal docs for its golden path cover only service creation: the scaffolding command, the generated directory layout, and a link to the component architecture. Any lifecycle question an engineer cannot answer from those docs becomes a ticket to the platform team. Last quarter's tickets, with the platform team's estimate of the platform-engineer minutes each ticket consumed and the share a written doc would have prevented: | Ticket category | Tickets | Minutes per ticket | Share a doc would prevent | | --- | --- | --- | --- | | Create a new service | 240 | 5 | 90% | | Rotate credentials or change config | 90 | 60 | 80% | | Diagnose a failed pipeline at 02:00 | 60 | 90 | 70% | | Migrate a service off the platform | 12 | 300 | 50% | | Find the service owner | 30 | 20 | 60% | The team can write one doc this quarter. Which one removes the most platform-engineer time?
MCQ-4A platform team maintains a service template that scaffolds a repository with a Dockerfile, CI pipeline, Helm chart, and SLO dashboard. Version 3.2 is current. Across 40 services, 12 still run a base image from template v2, 5 have hand-edited their CI pipeline, and 3 miss their SLO. The team proposes upgrading every service to template v4, which adds a newer base image and a revised dashboard. What does adopting v4 actually establish for these services?
MCQ-7A platform team publishes a golden-path service template to 180 services as a mutable `latest` reference. Each service's pipeline runs once every 10 days on average and picks up whatever `latest` points to. A breaking change broke 52 services over the next two weeks, and build logs recorded only `template: latest`, so nobody could tell which revision a service had built against. The team switches to immutable semver tags and records the resolved tag in each build's metadata; owners now bump deliberately, which historically happens for 6 services per week. For the next breaking change, roughly how long until half the fleet is exposed?
MCQ-9A platform team owns a service template that 240 repositories were generated from. The template's CI runs `golangci-lint` pinned to v1.55 plus `go test ./...`. The team bumps the pin to v1.60, generates 5 fresh repositories from the template, runs their CI, all green, then merges the template-sync PRs into all 240. Within a day, 40 of the 240 fail CI on a lint check that v1.60 ships and v1.55 did not, and every failure is on a line the sync PR did not touch. The 5 canary repositories still pass. Which change to the release plan would have caught this before the sync PRs merged?
MCQ-6A platform team's golden path template scaffolds a repo, a CI pipeline, a deploy job, and an ownership manifest naming the on-call team. Sixty services use it. Last quarter the template's median repo-to-first-deploy was 1.5 days and 92% of merges reached production without manual steps. Across those services there were 41 Sev-2 incidents with a median page-to-mitigation of 5h40m. In 28 of the 41, the responder identified the owning team within 5 minutes but spent more than half the mitigation time reconstructing how to revert the last deploy; no service scaffolded by the template ships a way to revert its last deploy. Which change to the template most directly reduces the median page-to-mitigation?
MCQ-15A platform team runs a self-service 'provision service' path. Last month 47 of 120 runs failed. In 38 of those, the user saw only 'provision failed' after 12 minutes; platform engineers then spent an average 35 minutes reading controller logs to classify each failure as a quota rejection, a bad manifest, or a transient API timeout. The path already retries transient API calls three times. Which change best lets users diagnose and recover from these failures themselves?
MCQ-21A platform team's scaffolder, the creation control plane that turns templates into new services, persists each task's steps in a task store as pending, running, or succeeded, and on restart a worker re-runs any step not marked succeeded. The push step commits the rendered tree with fixed commit metadata, so the same tree always produces the same commit. During a 40-minute deploy the workers die. 300 services are already running in production, and 12 template executions were in flight. Four of the 12 were in the 'push to git' step and had already pushed their commit when the worker died. What is the correct outcome?
MCQ-16A platform team is replacing a hand-rolled deploy script with a new pipeline across 240 services. Before cutover they inventory the repos: 96 commit generated code (protobuf stubs, vendored deps) that the new pipeline regenerates at build time; 14 of those 96 carry hand-edits on top of the generated files that the generator cannot rebuild; 22 services keep state in local volumes on the deploy host. The plan is one cutover weekend: freeze commits Friday, run the new pipeline over all 240 repos, delete the old scripts Monday. The team wants the first wave to cover as many repos as possible without losing any content the inventory shows is not reproducible. Which revision does that?
MCQ-20checkout-api was generated six months ago from the platform team's golden-path template. The template ships 2 replicas and no PodDisruptionBudget, and its README says '2 replicas, no PDB: size these for your SLO.' The platform team's contract covers the template, the pipeline, the dashboards and the alert definitions; the workload team owns the service's runtime behavior and its SLO target, 99.9% monthly (about 43 minutes). During a routine node drain both replicas land on the drained node and checkout-api is down for 12 minutes. What is the correct split of the fix?
MCQ-5A platform team's golden-path module provisions services into prod. The module has a `data_classification` variable that defaults to `internal`. It also reads a `classification` tag from the subnet. Last week a service processing restricted data was deployed to prod because the subnet in the new region lacked the tag, so the module used the default `internal`. The policy engine allows `internal` data in prod. The compliance team now requires that classification be declared by the service team, not inferred from network tags. Which change to the module satisfies that requirement and would have blocked the deployment?
MCQ-10A generator emits the same Deployment template for 40 services. Each pod reads its database password from `envFrom: secretRef: db-creds` at container start and opens a 20-connection pool. The password rotates every 24 hours in the external store, which keeps the previous password valid for 24 hours after each rotation. The ExternalSecret's `refreshInterval` is 1 hour, and you have confirmed that the Kubernetes Secret holds the new value within an hour of every rotation. Pods that have been running for more than 48 hours log `password authentication failed` on every new connection; restarting a pod clears the errors. What is the cause?
MCQ-12A payments team needs a canary rollout the golden-path chart doesn't support, so they ship their own Helm chart. It renders a Deployment with `requests: {cpu: 500m, memory: 512Mi}`, `limits: {cpu: 1, memory: 1Gi}`, the label `platform/owner: checkout`, and a PodDisruptionBudget with `maxUnavailable: 1`. The Kyverno policy on the `payments` namespace requires every Deployment to set CPU and memory requests and limits, carry `platform/owner`, and be covered by a PDB with `minAvailable >= 1`. What must they change for the release to be admitted?
MCQ-8A scaffolding service renders Go `text/template` files from a web form, then runs a post-generation hook: `sh -c "git add -A && git commit -m 'scaffold {{.ServiceName}}'"`. The pod's ServiceAccount is bound to a Role that allows `create` and `delete` on Deployments in the `prod` namespace, because the same pod applies and rolls back generated manifests. A developer submits `ServiceName` = `api; kubectl delete deploy payments -n prod`. The generated files are correct, but the payments Deployment is deleted. Which change prevents this submission from deleting the Deployment?
MCQ-11A payments team's CI gate blocks merges on any SAST finding. Last quarter it raised 1,400 findings across 900 pull requests; 82% were in generated code or test fixtures. Of the 279 PRs merged with --force, none contained a finding the security team later confirmed as real. Median time from block to security review was 2.5 days. The team wants to keep the gate blocking merges. Which change is most likely to cut the forced merges?
MCQ-13Your platform group runs a paved-road PostgreSQL 15 service with automated backups and IAM auth. A fraud-scoring team asks for an off-road exception to run PostgreSQL 13 on EC2 because their vendor appliance only certifies 13. The service processes 4,000 TPS and stores card tokens. The request names the vendor constraint and a migration ticket, but no owner, no compensating controls, no evidence of the vendor's certification timeline, and asks for an 18-month expiry. The review policy requires every exception to be bounded. What should the review board do?
MCQ-22A team is retiring /v1 of an internal pricing API. The gateway access log is sampled at 1 in 500. Over the last 30 days the sampled log holds 9 requests to /v1, all from client ID billing-batch, which the client registry lists as a nightly job owned by the data team. The team's checklist requires zero live consumers before the handler is deleted, and treats under 1 request per day as dead. What should the team do?
MCQ-24An API gateway team is canarying a new connection-pool config on the shared database proxy. The release bar is 1% failed requests in every cohort. The canary dashboard headlines 0.31% and the release is queued for promotion. The window holds 990,000 stateless-service requests with 1,980 failures and 10,000 stateful-service requests with 1,120 failures; the previous config ran stateless at 0.2% and stateful at 0.3%. What should the team do?