Skip to content
Tech Interview Prep
Technical interview guide

Internal Developer Platforms

The self-service layer that lets product engineers provision infrastructure without filing a ticket to a platform team.

Read
35 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 Capability Model, Google Cloud platform engineering guidance, and CNCF Platform Engineering Maturity Model current 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-7A platform team has 80 services. They run Backstage as a catalog and a 'Create service' button that opens a GitHub repo from a template. To get a namespace, database and CI credentials, developers still file a ticket; a platform engineer runs Terraform and kubectl by hand. A policy engine checks Terraform plans in CI but is not wired to Backstage. Leadership asks whether the company now has an internal developer platform. Which answer fits the situation?
MCQ-8A platform team's IDP takes a service manifest (image, replica count, SLO, data classification) and provisions the cluster, node pools, ingress and DNS behind it. Developers keep filing tickets asking which underlying resource their service landed on. Finance needs per-service spend, incident response needs a pageable owner, compliance needs the region constraint enforced, and on-call needs failure signals plus the scaling limits that stop one service exhausting the cluster. What should the abstraction hide?
MCQ-10A platform team owns a shared sidecar-injection template used by 180 services. A commit adds an init container that waits for a ConfigMap named `app-config` in each service namespace. The team applies the template to all 180 services in one GitOps sync. Within 10 minutes, 47 services fail readiness because `app-config` is missing. If the team had first applied the template to a canary of 10 services, what is the expected number of readiness failures in that canary, assuming the same failure rate?
MCQ-21A platform team's service template ships 24 files. Each repo keeps a `.template-manifest` recording the template version and the SHA-256 of every file as the generator wrote it. The upgrade tool's contract is that it replaces a file in place only when the file's current hash still matches the manifest; otherwise it writes the new version as `<file>.new` and leaves the original untouched. Template v4 changes 7 of the 24 files. In one repo, 3 of those 7 files no longer match the manifest. After the upgrade runs, how are the 7 changed files handled?
MCQ-16A platform team's admission policy rejects any manifest that omits a `resources.limits` block. Last quarter the policy produced 1,800 rejections. 71% of the rejected manifests were re-applied unchanged after an on-call platform engineer granted an exemption, with a median exemption turnaround of 2.5 days. The number of non-compliant workloads running in production was flat over the quarter. 84% of the rejected manifests came from the platform's own service template, which omits the limits block. Which change is most likely to reduce non-compliant workloads in production?
MCQ-14Your platform team is preparing its quarterly adoption slide. The new deploy pipeline is supported for the 400 engineers who own services in the migrated repositories; the company has 1,200 engineers in total. Over the last 90 days, 260 of those 400 owners deployed through the pipeline at least once, 1,044 of the 1,200 engineers logged into the platform portal at least once, and the portal lists 12 golden-path templates. The draft slide reads "Adoption: 87%". Which number and justification belong on the slide?
MCQ-24A self-service provisioning portal accepted 4,000 requests last quarter. It returns HTTP 202 as soon as a request is queued, and its API availability dashboard reads 99.98%. 480 of the requests later failed at the cloud provider's quota check, and of the rest, 15% were still not ready 10 minutes after the request, against a published 10-minute commitment. What provisioning success rate should the platform report for the quarter?
MCQ-19A platform team runs an internal provisioning API: a single `POST /queues` that takes an opaque `spec` string and returns `{"ok": true}` or `{"ok": false}`. Over one quarter it produced 276 failed requests. On-call's postmortem: | count | median on-call time | what happened | | --- | --- | --- | | 118 | 20 min | the developer passed a provider-specific option the platform doesn't support (e.g. `replicas: 3` on a queue with no replicas); the provider rejected it 90 s after the request was accepted | | 74 | 3 h | two teams each provisioned a queue for the same workload; neither could see the other's queue, so both ran until the bill arrived | | 48 | 15 min | the developer gave up waiting and re-submitted while the first request was still running | | 36 | 25 min | the failure response said only `{"ok": false}`; on-call re-ran the whole flow to find which step broke | The team can ship one interface change this quarter. Which one removes the most on-call time?
MCQ-1Your platform team of six supports 300 engineers and 90 services. The portal you shipped links to Jenkins, Grafana, the wiki, and a Jira form for infrastructure requests. Six months in, 62% of the 140 open Jira tickets are 'add a secret to service X' or 'grant deploy access to service Y', median time from repo creation to first production deploy is 11 days, and the backlog has grown every month. SOC 2 requires every production secret grant and deploy-permission change to record an approver identity. Which change best addresses the measured bottleneck?
MCQ-11A platform team's service template scaffolds a repo with a Dockerfile, a CI workflow that runs unit tests and builds an image in 11 minutes, and a README. Deploys are run by hand with kubectl apply. Of the 40 services created from it in the past year, 14 have no owner recorded in any system, 71% of Sev-2 incidents were first reported by a customer rather than by an alert, and median merge-to-production is 6 days. Which change to the template addresses the cause common to all three figures?
MCQ-4A platform team ships a database provisioning portal and calls it self-service. Last quarter it took 1,200 requests. The 940 for the three standard instance sizes were provisioned by Terraform in a median of 4 minutes, and the requester got status, a connection string and an audit record. The other 260 asked for non-standard sizes: the portal opened a Jira ticket for the DBA on-call, who provisioned by hand in a median of 3 days, and the requester saw only "submitted". Which change would most increase the share of requests that are genuinely self-service?
MCQ-2A platform team runs an internal developer portal. Of the org's 1,900 services, 340 (18%) deploy through the portal; the rest still go through the old Jenkins jobs or manual runbooks. The team shipped 14 features last quarter and the portal holds 99.98% uptime, but the developer satisfaction score is 3.1/5. Exit interviews with the 12 teams that tried the portal and went back to Jenkins all cite the same blocker: the portal's golden path assumes a containerized service, and their services are Java 8 WARs on EC2. The team has three engineers and one quarter. What should they do?
MCQ-5A platform team's golden path for new services is a repo template that wires CI, a pre-approved Helm chart pinned at 512 MiB memory and 3 replicas, an mTLS sidecar and a Grafana dashboard. Nine months after launch, 80 of 210 services use it. The 130 that don't are the same kind of request/response service as the adopters and differ only in size, because each serves a different tenant: memory needs run 256 MiB to 4 GiB, replicas 1 to 12, and the 130 fall into 61 distinct (memory, replicas) combinations, none shared by more than 7 services. Changing either value in the chart means a platform ticket that averages six days. The team has two engineers and one quarter. Which change is most likely to move adoption?
MCQ-6A platform team's golden path ships a 5 s HTTP timeout and a 2-replica default across 140 services. Its exception policy grants a 90-day registered deviation and promotes any deviation held by three or more teams into the next golden-path revision. At the Q2 review, seven teams hold the same exception: a 30 s timeout for calls to the legacy billing gateway, whose p99 is 22 s. What should the review produce?
MCQ-9An internal deploy platform serves 40 product teams. Adoption is 90% (36 teams), but 12 teams maintain their own wrapper scripts, and 11 of those 12 wrap the same thing: the release-approval step, which takes six API calls plus a manual ticket. Onboarding docs are the top support topic at 45 of 120 weekly tickets, and median time-to-first-deploy for a new team is 3 weeks. Next quarter's roadmap is already committed to three features requested by the two largest teams. Which move follows from running the platform as a product?
MCQ-18Your platform team shipped a scaffold-and-pipeline internal developer platform six weeks ago and leadership mandated it for all 60 teams on day one. Production deploys still run through the shared Jenkins instance. Six weeks in: 47 of 60 teams have a repo created from the template, 11 teams have routed any deploy through the platform pipeline (19% of production deploys), median time-to-first-deploy for a new service is still 9 days, 60% of incoming platform tickets say the template does not fit our Java 8 service, and two surveys have returned the same line: the template does not fit our service. You have two engineers and one quarter before the expansion decision goes to planning. What should the platform team do?
MCQ-25A 400-engineer org has run an internal developer platform (IDP) for 12 months. Across 620 services: 78% are on the golden path; median time from repo creation to first production deploy fell from 6 weeks to 3 days; platform-team "how do I…" tickets fell from 210 to 60 per month; change failure rate moved 18% → 16%; median deployment frequency 4 → 9 per week. 22% of services use a documented escape hatch. The two most common reasons are a missing Kafka-topic provisioning module and a build cache that does not support Bazel monorepos; two teams maintain their own CI for the same reason. The platform team has one quarter of capacity. Which investment does the data best support?
MCQ-22A 520-developer org set its platform team a half-year OKR to cut provisioning tickets by 50%. Tickets fell from 400 in January to 200 in June. Portal monthly active developers fell from 320 to 260 while headcount grew from 400 to 520. Provisioning events per 100 developers, counted in the audit logs across the portal and direct cloud accounts, were 45 in January and 46 in June. Two of the 20 teams also took a vendor managed-service exemption. What is the best-supported conclusion?
MCQ-3A platform team instrumented 120 backend engineers for a quarter. Median PR-to-merge is 3.2 days. Of that wall-clock time, 41% is engineers waiting on the single shared staging environment, which serializes one deploy at a time, and 12% is rebuilding local environments that drift from CI. The backlog's top request, from 38 engineers, is a self-service service catalog; telemetry shows 6% of engineers opened the existing catalog in the last 90 days. The team has one project's worth of budget and a goal of bringing median PR-to-merge under 2 days. Which action does the evidence support?
MCQ-20A platform team of four supports 180 developers and is deciding, component by component, what to build and what to buy. The build cache: a managed service quotes $0.002 per cache-restore request, and the team's CI makes 9 million of them a year. Self-hosting the same cache on existing cluster capacity would take one engineer twelve weeks to build, then about two hours a week to patch and operate. A platform engineer costs $4,000 a week fully loaded, and the team prices its own on-call time at $100 an hour. The release-approval workflow: it encodes a sign-off matrix that the compliance team revises roughly every quarter. Both vendors evaluated hard-code a fixed approver list. The last time a fixed list met a rule change, releases froze for three days. Which decision fits both components?
MCQ-23A payments team must migrate 4 TB of transaction history. The paved-road managed queue caps messages at 256 KB, but the records are single 5 MB payloads that cannot be split, so the team asks to run its own RabbitMQ on the shared Kubernetes cluster for the 12 weeks the migration takes. Platform on-call has no RabbitMQ runbooks and a 15-minute response SLA; the payments team has its own SRE. Security requires every catalog deviation to have a named owner on record. What should the platform team do?
MCQ-12A provisioning service creates cloud instances. Its worker calls the provider's create API, which returns 202 Accepted with an operation ID; the instance becomes ready about 90 seconds later. The worker writes status 'succeeded' as soon as it sees the 202 and exits, and job state lives only in the worker's memory. After a rolling restart, 47 of 1,200 jobs read 'succeeded' but their instances never received DNS records, and re-running those jobs created a second instance each. Which change removes both symptoms?
MCQ-15A platform control plane reconciles 12,000 cloud resources against Git manifests every 60 seconds. Its operating requirements are: an unreviewed change must never become the declared state; the owning team must be told when a change is reverted; drift must not survive past the next reconciliation pass; and one tenant's drift must not block another tenant's deploys. What should the control plane do when it detects drift?
MCQ-13Acme's internal developer platform runs on a management cluster: a Backstage portal, an Argo CD controller, and a validating admission webhook (failurePolicy: Fail, timeoutSeconds: 10) that all 9 workload clusters call to admit new pods. At 02:00 the management cluster's API server crashes. By 02:05 the on-call sees 412 application pods still serving traffic across the 9 clusters, 37 Argo syncs stuck in Pending, and a Sev-1 from a team that cannot deploy. No workload cluster reports pod evictions. What is the correct read of this outage and the right first action?
MCQ-17A platform team's IDP control plane provisions 40 Kubernetes clusters (12 production, 28 dev). A single long-lived provisioning runner holds one cluster-admin kubeconfig that reaches all 40 clusters, and it renders templates by executing plugin code in-process. A developer with write access to the shared template repo merges a plugin whose post-provision hook reads the runner's environment and mounted files. The template is instantiated in 3 dev clusters. If the hook is malicious, what does it obtain?