Design CI/CD for 120 microservices, 600 production deploys per day, and a ten-minute commit-to-production target.

I would use reusable pipeline templates with independent service builds and build each commit only once. - Pull requests run affected lint, unit, and policy checks in under five minutes using dependency-aware caching. - Merge builds an immutable image, scans it with Trivy, signs it with Cosign, and promotes the same digest through environments. - Production uses automated canaries with SLO gates, while runner concurrency and deployment queues are sized for the measured 600-deploy peak.

A monorepo contains 250 services, but a one-line change triggers a 45-minute full build. How would you redesign CI?

I would build only the affected dependency graph and make cache keys reproducible from inputs. - Bazel, Nx, or Pants can map changed files to services and include reverse dependencies that need retesting. - Remote caches store compiled layers and test outputs by source, lockfile, toolchain, and build configuration hash. - A nightly full build catches graph mistakes, while pull-request p95 must fall below ten minutes without reducing required tests.

Your CI platform runs 1,000 jobs per hour, each averaging six minutes. How would you size and isolate runners?

I would provision roughly 150 concurrent ephemeral runners, above the 100-runner average needed for that load. - Autoscaling keeps about 20% warm capacity so queue p95 stays below two minutes during bursts. - Untrusted pull requests run without cloud credentials on separate node pools or accounts from signed release jobs. - Every job gets a clean VM or Pod, a 30-minute maximum lifetime, restricted egress, and no shared Docker socket.

Design a blue-green deployment for an API serving 20,000 requests per second with a ten-minute rollback target.

I would run green at full production capacity before shifting traffic, accepting temporary 2x compute cost. - Green receives smoke and synthetic tests, then 5% shadow or live traffic while latency, errors, and saturation are compared with blue. - The load balancer moves traffic in controlled steps and keeps connection draining below the ten-minute rollback window. - Blue remains deployable and connected to a backward-compatible data layer until green holds SLOs for at least 30 minutes.

Define a canary rollout for a service with 50 replicas, 99.9% availability, and a normal error rate of 0.2%.

I would progress through 5%, 25%, 50%, and 100% traffic only while canary signals stay close to baseline. - Each stage runs at least ten minutes and compares error rate, p95 latency, saturation, and a business metric such as checkout success. - Automatic rollback triggers above 0.5% errors, a 20% latency regression, or two fast burn-rate windows. - At least three canary replicas avoid judging one noisy instance, and the rollout pauses if sample volume is too small.

A service deploys 30 times per day and needs a non-null database column without downtime. How does the pipeline ship it?

I would use an expand-and-contract sequence split across at least three independently deployable changes. - First add the nullable column or compatible default without a long table rewrite, then deploy code that writes both representations. - Backfill in bounded batches while monitoring lock time, replica lag, and database CPU below agreed limits. - Only after all readers use the new field does a later pipeline validate completeness, add the constraint, and remove the old path.

8 environments currently rebuild the same application image. How would you enforce build-once, promote-many?

I would create one signed image digest after merge and promote metadata that references that digest. - Tests, SBOM, provenance, and vulnerability results attach to the immutable digest in the registry. - Environment-specific configuration stays outside the image in Helm values, Kustomize overlays, or runtime configuration. - Admission rejects mutable tags in production, so all eight environments run the artifact that passed the earlier gates.

Production rollback must finish within five minutes across 40 services. What must the delivery architecture provide?

I would make rollback a precomputed deployment action, not a rebuild or a manual manifest edit. - The registry retains recent signed digests, and GitOps history records the exact image and configuration revision for every release. - Database changes remain backward-compatible for at least one release window, otherwise application rollback is unsafe. - A quarterly exercise rolls back representative services and measures traffic recovery below five minutes, including readiness and connection draining.

A pipeline takes 28 minutes: 6 for build, 18 for tests, and 4 for deploy. How would you reach a 12-minute target?

I would attack the 18-minute test path first because it dominates the critical path. - Tests are split by historical duration across parallel workers, while flaky tests are quarantined with owners rather than blindly retried. - BuildKit layer caching and prebuilt toolchain images reduce the six-minute build without weakening reproducibility. - Deploy starts only after required gates, and p95 duration by stage proves whether the total stays below 12 minutes for two weeks.

You must publish amd64 and arm64 images for 70 services without doubling a 15-minute build. What would you change?

I would use native builders for each architecture and combine their manifests into one multi-architecture digest. - BuildKit executes architecture-independent stages once where possible and shares remote cache layers across runner pools. - Native arm64 runners avoid the large and variable emulation cost of QEMU on compile-heavy services. - CI runs smoke tests on both images and accepts the design only if p95 build time stays under 18 minutes.

CI receives 400 commits per hour and each pipeline uses eight runner-minutes. How much concurrency would you provision?

The average demand is about 54 concurrent runners, so I would start near 80 to absorb bursts and maintenance. - The calculation is 400 times eight divided by 60, then headroom is validated against arrival peaks rather than a flat hourly average. - Autoscaling uses queue depth and oldest-job age, with a two-minute p95 queue target. - Per-repository concurrency and cancellation of superseded commits stop one busy branch from consuming all 80 runners.

Two hundred production deploys run daily, but security wants a human approval on every one. How would you design the gate?

I would use risk-based automated gates and reserve human approval for high-impact changes. - Signed artifacts with passing tests, policy checks, and canary evidence can auto-promote for low-risk services. - IAM, database, public-network, or tier-0 changes require two named approvers and an expiring deployment window. - The platform records approver, digest, policy result, and rollback link, targeting under ten minutes median wait without weakening tier-0 control.

Eighty AWS accounts use Terraform, and teams have copied the same VPC code 35 times. How would you structure modules?

I would publish small versioned modules and compose them into supported account and environment stacks. - A VPC module owns networking primitives and outputs a stable contract, but does not also create EKS, databases, and application IAM. - Opinionated wrapper modules provide approved defaults for the common path while exposing only inputs teams genuinely vary. - CI tests examples, upgrade paths, policy compliance, and plans against at least two real account classes before release.

A single Terraform state contains 50,000 resources and takes 25 minutes to plan. How would you split it?

I would split state by lifecycle, blast radius, ownership, and change frequency rather than by arbitrary resource count. - Network, shared identity, clusters, data services, and application stacks become separate states with explicit remote outputs. - I avoid long chains of terraform_remote_state by publishing stable values through a parameter store or platform API where appropriate. - Migration uses moved blocks or state moves in reviewed batches, with lock time and plan p95 targeted below five minutes.

Design the Terraform backend for 40 teams applying up to 100 changes per day.

I would use a remote encrypted backend with locking, versioning, audit logs, and narrowly scoped identities. - S3 native lockfiles with versioning or Terraform Cloud prevent concurrent writes and retain recoverable state versions. - Each state has a dedicated apply identity; plan jobs receive read access and short-lived OIDC credentials. - CI serializes applies per state, times out stale locks through an audited procedure, and tests restoration quarterly.

Thirty accounts each have dev, staging, and production. Would you use Terraform workspaces or separate root configurations?

I would use separate root configurations and state backends for account and production boundaries, not 90 CLI workspaces. - Directory or generated-stack structure makes account, region, and environment visible in review and access policy. - Production gets a distinct apply identity and approval path, so selecting the wrong workspace cannot target it. - Workspaces remain useful for short-lived, identical preview environments where isolation and configuration truly match.

You manage 12,000 Terraform resources, and operators still make emergency console changes. How would you detect drift?

I would run scheduled read-only plans per state and route actionable drift to the owning team. - High-risk states plan hourly, ordinary states daily, and cloud audit events trigger targeted plans after console changes. - The report separates provider normalization from real create, update, and delete actions to keep false positives below 5%. - Emergency changes carry an owner and expiry, then are either imported into code or reverted within 24 hours.

A shared Terraform module has 60 consumers and needs a breaking input change. How would you release it?

I would publish a new major version and keep the old contract supported through a measured migration window. - Upgrade tests run representative consumer plans and document state moves, renamed inputs, and provider-version requirements. - Renovate or a similar bot opens pinned-version pull requests in cohorts of five consumers, starting with non-production. - The old major is removed only after all 60 states migrate and rollback instructions have been exercised once.

How would you separate Terraform plan and apply for 300 infrastructure pull requests per month?

I would generate the plan from the reviewed commit and apply that exact approved plan with a short-lived production identity. - CI stores the binary plan, human-readable summary, provider lockfile, and commit SHA as one immutable review artifact. - Apply refuses a changed branch head, stale plan, policy failure, or expired approval instead of silently replanning. - Production plans expire after a short window such as four hours because external state may change underneath them.

A company has 2,000 manually created cloud resources and wants Terraform ownership within six months. What is the migration design?

I would migrate by service boundary and risk tier, not attempt one bulk import. - Inventory and cloud tags establish owner, dependencies, and whether each resource should be imported, replaced, or retired. - Generated import blocks accelerate capture, but every imported state must reach a zero-change plan before normal applies begin. - Cohorts start with non-production and target about 100 resources per week while monitoring unintended replacement and access changes.

DevOps Engineer interview questions

100 real questions with model answers and explanations for Senior candidates.

See a DevOps Engineer resume example →

Practice with flashcards

Spaced repetition · Hunter Pass

Questions

microservicesdesignci-cd

I would use reusable pipeline templates with independent service builds and build each commit only once.

Pull requests run affected lint, unit, and policy checks in under five minutes using dependency-aware caching.
Merge builds an immutable image, scans it with Trivy, signs it with Cosign, and promotes the same digest through environments.
Production uses automated canaries with SLO gates, while runner concurrency and deployment queues are sized for the measured 600-deploy peak.

Why interviewers ask this: The interviewer checks whether the candidate can turn delivery volume and latency into a secure, reusable pipeline architecture.

monorepo

I would build only the affected dependency graph and make cache keys reproducible from inputs.

Bazel, Nx, or Pants can map changed files to services and include reverse dependencies that need retesting.
Remote caches store compiled layers and test outputs by source, lockfile, toolchain, and build configuration hash.
A nightly full build catches graph mistakes, while pull-request p95 must fall below ten minutes without reducing required tests.

Why interviewers ask this: A strong answer combines change detection, correct cache invalidation, and a full-build safety net.

I would provision roughly 150 concurrent ephemeral runners, above the 100-runner average needed for that load.

Autoscaling keeps about 20% warm capacity so queue p95 stays below two minutes during bursts.
Untrusted pull requests run without cloud credentials on separate node pools or accounts from signed release jobs.
Every job gets a clean VM or Pod, a 30-minute maximum lifetime, restricted egress, and no shared Docker socket.

Why interviewers ask this: The interviewer evaluates concurrency math, burst headroom, and security separation between trusted and untrusted jobs.

designdeployment-strategiesdeployment

I would run green at full production capacity before shifting traffic, accepting temporary 2x compute cost.

Green receives smoke and synthetic tests, then 5% shadow or live traffic while latency, errors, and saturation are compared with blue.
The load balancer moves traffic in controlled steps and keeps connection draining below the ten-minute rollback window.
Blue remains deployable and connected to a backward-compatible data layer until green holds SLOs for at least 30 minutes.

Why interviewers ask this: A strong answer makes capacity duplication, traffic control, data compatibility, and rollback timing explicit.

replicationdeployment-strategies

I would progress through 5%, 25%, 50%, and 100% traffic only while canary signals stay close to baseline.

Each stage runs at least ten minutes and compares error rate, p95 latency, saturation, and a business metric such as checkout success.
Automatic rollback triggers above 0.5% errors, a 20% latency regression, or two fast burn-rate windows.
At least three canary replicas avoid judging one noisy instance, and the rollout pauses if sample volume is too small.

Why interviewers ask this: The interviewer checks whether canary stages, sample size, and rollback gates are numerically defensible.

databaseschemadeployment

I would use an expand-and-contract sequence split across at least three independently deployable changes.

First add the nullable column or compatible default without a long table rewrite, then deploy code that writes both representations.
Backfill in bounded batches while monitoring lock time, replica lag, and database CPU below agreed limits.
Only after all readers use the new field does a later pipeline validate completeness, add the constraint, and remove the old path.

Why interviewers ask this: A strong answer integrates database compatibility and measured backfill safety into continuous delivery.

I would create one signed image digest after merge and promote metadata that references that digest.

Tests, SBOM, provenance, and vulnerability results attach to the immutable digest in the registry.
Environment-specific configuration stays outside the image in Helm values, Kustomize overlays, or runtime configuration.
Admission rejects mutable tags in production, so all eight environments run the artifact that passed the earlier gates.

Why interviewers ask this: The interviewer evaluates artifact immutability and separation of deployable code from environment configuration.

rollback

I would make rollback a precomputed deployment action, not a rebuild or a manual manifest edit.

The registry retains recent signed digests, and GitOps history records the exact image and configuration revision for every release.
Database changes remain backward-compatible for at least one release window, otherwise application rollback is unsafe.
A quarterly exercise rolls back representative services and measures traffic recovery below five minutes, including readiness and connection draining.

Why interviewers ask this: A strong answer treats rollback as an artifact, configuration, and data-compatibility capability that must be timed.

deploymentci-cdtesting

I would attack the 18-minute test path first because it dominates the critical path.

Tests are split by historical duration across parallel workers, while flaky tests are quarantined with owners rather than blindly retried.
BuildKit layer caching and prebuilt toolchain images reduce the six-minute build without weakening reproducibility.
Deploy starts only after required gates, and p95 duration by stage proves whether the total stays below 12 minutes for two weeks.

Why interviewers ask this: The interviewer checks whether optimization follows measured critical-path time rather than generic pipeline tuning.

I would use native builders for each architecture and combine their manifests into one multi-architecture digest.

BuildKit executes architecture-independent stages once where possible and shares remote cache layers across runner pools.
Native arm64 runners avoid the large and variable emulation cost of QEMU on compile-heavy services.
CI runs smoke tests on both images and accepts the design only if p95 build time stays under 18 minutes.

Why interviewers ask this: A strong answer balances multi-architecture correctness with native execution and reusable build layers.

ci-cdconcurrency

The average demand is about 54 concurrent runners, so I would start near 80 to absorb bursts and maintenance.

The calculation is 400 times eight divided by 60, then headroom is validated against arrival peaks rather than a flat hourly average.
Autoscaling uses queue depth and oldest-job age, with a two-minute p95 queue target.
Per-repository concurrency and cancellation of superseded commits stop one busy branch from consuming all 80 runners.

Why interviewers ask this: The interviewer evaluates basic capacity math plus fair scheduling and burst controls for a shared CI fleet.

designdeployment

I would use risk-based automated gates and reserve human approval for high-impact changes.

Signed artifacts with passing tests, policy checks, and canary evidence can auto-promote for low-risk services.
IAM, database, public-network, or tier-0 changes require two named approvers and an expiring deployment window.
The platform records approver, digest, policy result, and rollback link, targeting under ten minutes median wait without weakening tier-0 control.

Why interviewers ask this: A strong answer scales governance by change risk and preserves an auditable production path.

networkingterraform

I would publish small versioned modules and compose them into supported account and environment stacks.

A VPC module owns networking primitives and outputs a stable contract, but does not also create EKS, databases, and application IAM.
Opinionated wrapper modules provide approved defaults for the common path while exposing only inputs teams genuinely vary.
CI tests examples, upgrade paths, policy compliance, and plans against at least two real account classes before release.

Why interviewers ask this: The interviewer checks module boundaries, supported defaults, and evidence that reusable IaC can evolve safely.

terraform

I would split state by lifecycle, blast radius, ownership, and change frequency rather than by arbitrary resource count.

Network, shared identity, clusters, data services, and application stacks become separate states with explicit remote outputs.
I avoid long chains of terraform_remote_state by publishing stable values through a parameter store or platform API where appropriate.
Migration uses moved blocks or state moves in reviewed batches, with lock time and plan p95 targeted below five minutes.

Why interviewers ask this: A strong answer improves plan scale without replacing one monolith with a fragile graph of state dependencies.

designterraform

I would use a remote encrypted backend with locking, versioning, audit logs, and narrowly scoped identities.

S3 native lockfiles with versioning or Terraform Cloud prevent concurrent writes and retain recoverable state versions.
Each state has a dedicated apply identity; plan jobs receive read access and short-lived OIDC credentials.
CI serializes applies per state, times out stale locks through an audited procedure, and tests restoration quarterly.

Why interviewers ask this: The interviewer evaluates state durability, concurrency control, identity scope, and recoverability under real change volume.

terraformconfig

I would use separate root configurations and state backends for account and production boundaries, not 90 CLI workspaces.

Directory or generated-stack structure makes account, region, and environment visible in review and access policy.
Production gets a distinct apply identity and approval path, so selecting the wrong workspace cannot target it.
Workspaces remain useful for short-lived, identical preview environments where isolation and configuration truly match.

Why interviewers ask this: A strong answer chooses visible, access-controlled isolation for long-lived environments and limits workspaces to homogeneous copies.

terraformiac

I would run scheduled read-only plans per state and route actionable drift to the owning team.

High-risk states plan hourly, ordinary states daily, and cloud audit events trigger targeted plans after console changes.
The report separates provider normalization from real create, update, and delete actions to keep false positives below 5%.
Emergency changes carry an owner and expiry, then are either imported into code or reverted within 24 hours.

Why interviewers ask this: The interviewer checks whether drift detection is frequent, attributable, low-noise, and connected to a remediation path.

terraform

I would publish a new major version and keep the old contract supported through a measured migration window.

Upgrade tests run representative consumer plans and document state moves, renamed inputs, and provider-version requirements.
Renovate or a similar bot opens pinned-version pull requests in cohorts of five consumers, starting with non-production.
The old major is removed only after all 60 states migrate and rollback instructions have been exercised once.

Why interviewers ask this: A strong answer treats a module as a versioned product with consumer testing and controlled migration.

code-reviewterraform

I would generate the plan from the reviewed commit and apply that exact approved plan with a short-lived production identity.

CI stores the binary plan, human-readable summary, provider lockfile, and commit SHA as one immutable review artifact.
Apply refuses a changed branch head, stale plan, policy failure, or expired approval instead of silently replanning.
Production plans expire after a short window such as four hours because external state may change underneath them.

Why interviewers ask this: The interviewer evaluates review integrity and prevention of applying infrastructure different from what was approved.

ownershipmigrationsdesign

I would migrate by service boundary and risk tier, not attempt one bulk import.

Inventory and cloud tags establish owner, dependencies, and whether each resource should be imported, replaced, or retired.
Generated import blocks accelerate capture, but every imported state must reach a zero-change plan before normal applies begin.
Cohorts start with non-production and target about 100 resources per week while monitoring unintended replacement and access changes.

Why interviewers ask this: A strong answer combines inventory, import verification, and a sustainable migration rate for existing infrastructure.

Locked questions

21
Forty teams know TypeScript, but the platform group operates 15,000 Terraform resources. Would you add CDK or Pulumi?
terraformiactypescript
22
Security requires every new cloud database to be encrypted, private, and backed up for 30 days. Where do you enforce it in IaC?
databaseencryptioniac
23
Design Prometheus monitoring for 30 Kubernetes clusters producing 12 million active series with 13 months of retention.
kubernetesmonitoringdesign
24
Your platform ingests 20 TB of logs per day from 800 services. How would you design storage and querying?
queriesdesign
25
A tracing backend cannot store all 2 million spans per second. What sampling architecture would you use?
samplingarchitecture
26
How would you deploy OpenTelemetry collectors for 500 services across 20 clusters without one global bottleneck?
deploymentobservabilitytracking
27
Define SLIs for a checkout API handling 5,000 requests per second with a 99.95% availability target.
api
28
A service has a 99.9% monthly availability SLO. How much error budget does it have, and how would releases consume it?
reliabilityslo
29
Design burn-rate alerts for a 99.9% SLO without paging on every short error spike.
designsloalerting
30
Prometheus grows from 4 million to 18 million series after teams add user_id labels. What guardrails would you design?
guardrailsdesignmonitoring
31
Metrics, logs, and traces cost $140,000 per month. How would you set retention without weakening a 90-day audit requirement?
retentionmonitoring
32
A release dashboard compares 2 versions but cannot connect errors to traces or the deployed digest. How would you redesign telemetry correlation?
correlationdeployment
33
You are building an internal developer platform for 300 engineers and have 90 days for the first release. What do you ship?
34
Design a golden path for 80 Go services that currently take five days to reach production.
design
35
A 500-engineer organization wants Backstage, but its service catalog is only 40% accurate. What would you implement first?
36
Teams request 50 development databases per day. How would you provide self-service without giving them unrestricted cloud access?
database
37
Design a Crossplane platform API for three clouds without exposing 200 provider fields to application teams.
designapiiac
38
How would you measure an internal platform used by 45 teams instead of counting shipped features?
39
A product must run on AWS and GCP, but 85% of workloads use only basic compute, database, and object storage. What should the platform abstract?
database
40
GitHub Actions deploys to 25 AWS accounts using year-long access keys. Design the OIDC replacement.
ci-cddeploymentdesign
41
Two hundred services need database credentials rotated every hour. How would Vault deliver them?
databasesecrets
42
ArgoCD manages 20 clusters, and security forbids plaintext secrets in Git. Would you choose SOPS or External Secrets?
gitgitopssecrets
43
Design software supply-chain controls for 300 container builds per day without adding more than three minutes to CI.
designcontainers
44
Public pull requests need CI, but release runners can deploy to 12 production accounts. How do you prevent credential theft?
code-reviewdeployment
45
Design ArgoCD repository structure for 100 teams, 20 clusters, and four environments without one root application controlling everything.
designgitops
46
A release must move through 3 environments while preserving one image digest. How would GitOps promotion work?
gitops
47
An application needs a database migration, Deployment, and smoke test in a 15-minute ArgoCD sync. How do you order it?
databasemigrationsdeployment
48
ArgoCD detects 300 drifted resources per day, but 70% are fields mutated by other controllers. How would you configure sync safely?
configiacgitops
49
A container registry holds 200 TB, grows 12 TB monthly, and must preserve every production artifact for one year. What retention policy would you use?
containersartifactsregistries
50
Teams in three regions pull 8 TB of images daily from one registry, and deployments must work during a regional outage. What architecture would you choose?
deploymentregistriesarchitecture
51
A canary passed at 5% traffic, but after promotion checkout errors rose from 0.2% to 3%. What do you roll back and inspect?
deployment-strategiesrollback
52
A release keeps error rate flat but raises API p99 from 240 ms to 1.8 seconds at 50% traffic. Do you promote?
api
53
Deployments fail because a database migration takes 22 minutes while the pipeline timeout is 15. How do you recover?
databasemigrationsdeployment
54
A feature flag enables new code for 30% of users, and errors jump to 6% without a new deploy. What is your rollback path?
deploymentfeature-flagsrollback
55
Build time for a monorepo rises from 11 to 46 minutes in one week. How do you find the regression?
monorepo
56
CI queue p95 grows from 2 to 35 minutes during business hours, but runner CPU averages only 40%. What do you inspect?
data-structures
57
A remote build cache hit rate drops from 82% to 9% after a toolchain update. How do you restore it safely?
caching
58
Flaky tests fail 18% of pipelines, and teams rerun jobs until they pass. What do you change this week?
flakyci-cd
59
Artifact uploads add 14 minutes to every build after image size grows from 600 MB to 4 GB. What do you do?
artifacts
60
A runner image update breaks 70% of pipelines across 50 teams. How do you restore service?
ci-cd
61
A manually opened security group creates a public database path for 47 minutes. Terraform does not detect it until the next day. What do you do?
databaseterraformnetworking
62
Terraform plans replacement of 400 production resources after a provider upgrade. What is your decision process?
terraformconcurrency
63
A Terraform apply loses its lock after creating 60 of 100 resources. Another pipeline is waiting. What do you do?
terraformci-cd
64
A refactor moves Terraform resources between states, and production loses 12 IAM bindings. How do you recover?
refactoringterraform
65
A drift job reports 3,000 changes, but 95% are provider-generated ordering noise. How do you recover signal?
iac
66
A CI token with admin access to 18 AWS accounts is published in a public repository for 26 minutes. What happens first?
tokens
67
A database password is found in 240 commits and 35 release artifacts. Deleting the latest file is not enough. What do you do?
databasepasswordsartifacts
68
A GitHub OIDC trust policy accidentally allows every repository in the organization to assume a production role. How do you contain it?
69
Vault is unavailable for 25 minutes, and credentials for 80 services begin expiring. How do you keep services running safely?
secrets
70
Trivy blocks 65% of builds after a vulnerability database update, including 900 findings with no available fix. What do you change?
databasevulnerabilities
71
During a 40-minute checkout outage, you have CPU and error metrics but no queue-depth metric. How do you close the observability gap?
observabilitymonitoringdata-structures
72
API p99 triples after a deploy, but only 4% of requests have traces and none cover the slow endpoint. What do you fix?
endpointsdeployment
73
A logging pipeline drops 18% of production logs during peak traffic without paging anyone. How do you restore trust?
loggingci-cd
74
Prometheus query p95 rises to 28 seconds during incidents, making dashboards unusable. What do you change first?
incidentsqueriesmonitoring
75
The platform sends 600 alerts per day, but only 14 require action. What do you remove first?
alerting
76
Self-hosted runner cost jumps from $70,000 to $190,000 per month while job volume grows only 8%. How do you investigate?
ci-cd
77
A CI job downloads 6 GB of dependencies 20,000 times per month, creating a $38,000 network bill. What do you change?
dependencies
78
Observability spend grows from $110,000 to $260,000 per month after three teams add high-cardinality labels. What do you do?
observability
79
Log storage reaches 4 PB and grows 200 TB per month, but searches older than 14 days represent 0.3% of queries. What do you change?
queries
80
A release creates 900 idle GPU instances overnight and adds $54,000 before detection. How do you contain and prevent it?
81
At 02:00 the artifact registry returns 503 for 70% of pulls, blocking all new Pods. You are platform on-call. What do you do?
artifactsregistrieson-call
82
At 03:00 Terraform Cloud cannot start applies, and a network fix is waiting for production. How do you decide on break-glass?
terraform
83
A platform API used by 45 teams returns 30% errors after its certificate rotates. How do you lead the on-call response?
on-callapi
84
The error budget is exhausted, but marketing wants a launch expected to triple traffic tomorrow. What recommendation do you make?
reliability
85
A chaos test terminates one availability zone and recovery takes 19 minutes against a five-minute objective. What happens next?
availability
86
Traffic forecasting predicted 30% growth, but a partner launch creates 4x load and exhausts capacity in 12 minutes. What do you change?
capacity
87
The platform team spends 140 hours per month manually provisioning access that takes 25 minutes per request. What do you automate?
88
A Jenkins migration moves 80 of 800 jobs, but failure rate rises from 2% to 14%. Do you continue?
migrations
89
During VM-to-Kubernetes cutover, 10% traffic produces 4% errors while the VM path stays at 0.3%. What do you do?
kubernetes
90
A 2 TB PostgreSQL cross-cloud migration reaches cutover with 18 minutes of replica lag against a five-minute window. Do you proceed?
postgresmigrationsreplication
91
An ArgoCD sync unexpectedly deletes 120 resources because an ApplicationSet generator returns an empty list. What do you do?
gitopsgenerators
92
An ArgoCD PreSync hook hangs for 20 minutes and blocks 60 application releases. How do you recover?
hooksgitops
93
A critical CVE appears four hours before release, affects 300 images, and has no patch. What standard do you apply?
94
Fifty teams have unique pipelines, and a shared security fix reaches only 12 after two weeks. How do you standardize now?
ci-cd
95
A strong operations engineer resolves 20 pages per month but repeats the same manual fix. How do you mentor them?
mentoring
96
A mid-level engineer wants to apply a Terraform plan with 30 replacements to meet a deadline. How do you coach the decision?
estimationterraform
97
A six-person team has one Vault expert, and that person handled 9 of the last 10 secret incidents. What do you change this quarter?
secretsincidents
98
During a sev1 affecting 40% of customers, an executive posts fixes every two minutes and asks for constant updates. How do you handle it?
soft-skills
99
The CFO requires a 40% infrastructure cut, but 99.95% SLO services already use 70% of peak capacity. What do you propose?
capacityslo
100
An AI-generated Terraform pull request proposes 180 resources and passes syntax checks, but 12 IAM policies contain wildcards. What standard do you enforce?
code-reviewterraform

Questions

1Design CI/CD for 120 microservices, 600 production deploys per day, and a ten-minute commit-to-production target.

2A monorepo contains 250 services, but a one-line change triggers a 45-minute full build. How would you redesign CI?

3Your CI platform runs 1,000 jobs per hour, each averaging six minutes. How would you size and isolate runners?

4Design a blue-green deployment for an API serving 20,000 requests per second with a ten-minute rollback target.

5Define a canary rollout for a service with 50 replicas, 99.9% availability, and a normal error rate of 0.2%.

6A service deploys 30 times per day and needs a non-null database column without downtime. How does the pipeline ship it?

78 environments currently rebuild the same application image. How would you enforce build-once, promote-many?

8Production rollback must finish within five minutes across 40 services. What must the delivery architecture provide?

9A pipeline takes 28 minutes: 6 for build, 18 for tests, and 4 for deploy. How would you reach a 12-minute target?

10You must publish amd64 and arm64 images for 70 services without doubling a 15-minute build. What would you change?

11CI receives 400 commits per hour and each pipeline uses eight runner-minutes. How much concurrency would you provision?

12Two hundred production deploys run daily, but security wants a human approval on every one. How would you design the gate?

13Eighty AWS accounts use Terraform, and teams have copied the same VPC code 35 times. How would you structure modules?

14A single Terraform state contains 50,000 resources and takes 25 minutes to plan. How would you split it?

15Design the Terraform backend for 40 teams applying up to 100 changes per day.

16Thirty accounts each have dev, staging, and production. Would you use Terraform workspaces or separate root configurations?

17You manage 12,000 Terraform resources, and operators still make emergency console changes. How would you detect drift?

18A shared Terraform module has 60 consumers and needs a breaking input change. How would you release it?

19How would you separate Terraform plan and apply for 300 infrastructure pull requests per month?

20A company has 2,000 manually created cloud resources and wants Terraform ownership within six months. What is the migration design?

Locked questions