CDL practice questions and answers

Correct answer: D — Enable Autoclass on the bucket.

Autoclass moves each object between Standard, Nearline, Coldline, and Archive based on that object's own access pattern, and it moves objects back to Standard when they are read again. There are no early deletion charges and no retrieval charges for Autoclass transitions, and there is nothing to tune, which is exactly what an unpredictable pattern needs. The lifecycle rule in option A is a fixed guess: the files that are read months later are penalized with retrieval charges, and the rule never learns. Archive-only makes every read slow and costly. Staying on Standard keeps the bill high and only moves the problem to a quarterly review.

Correct answer: B — Rehost the application on Compute Engine virtual machines as it is, then modernize after the move

Rehosting, also called lift and shift, moves the workload with the fewest changes and is the approach that reliably meets a hard deadline; the company can modernize once it is out of the data center. Refactoring into microservices is the slowest path and needs deep knowledge of a codebase whose authors have left. Replatforming onto App Engine standard means rewriting for a supported runtime, which is more change than five months allows for a monolith. Rebuilding on Cloud Run is a full redesign and carries the highest risk of missing the exit date.

Correct answer: D — GKE Enterprise, to manage the on-premises and multicloud clusters from one control plane.

GKE Enterprise is built for exactly this: it registers clusters running on-premises, in Google Cloud, and in other clouds into one fleet, and applies consistent configuration and policy across all of them. Cloud Run does not manage the company's existing Kubernetes clusters. Autopilot only applies to clusters inside Google Cloud, so it cannot cover the data center or the second cloud. Migrate to Containers moves virtual machine workloads into containers; it does not govern clusters that already exist.

Correct answer: C — Siloed data, because customer information is split across department-owned systems with no shared definition

The symptoms named in the scenario are slow access to data and numbers that disagree between teams. Both point to data silos: each department owns its own copy and its own meaning of a customer. A is ruled out because the scenario says the engineers are trained. D is ruled out because the software is current, so technical debt is not the constraint here. B may be true in the market, but competitive pressure is a driver that pushes the bank to change, not the blocker stopping it.

Correct answer: B — Security Command Center, to scan the organization and report misconfigurations and vulnerabilities.

Security Command Center is the central security and risk platform for a Google Cloud organization: it inventories assets across all projects and raises findings for misconfigurations such as public buckets and over-open firewall rules. Cloud Audit Logs tell you what was done after the fact, but do not evaluate a configuration against best practice. Cloud Armor protects applications from incoming web attacks, and Cloud KMS manages encryption keys; neither one scans projects for misconfiguration.

Correct answer: C — Platform as a Service (PaaS), where the team deploys code and the provider runs the OS and runtime

PaaS is the model where you hand over application code and the provider operates the operating system, runtime, and scaling underneath. IaaS still leaves guest OS patching and machine management with the team, which is exactly the work they cannot do. SaaS delivers a ready-made application, so there is no place to deploy custom code. Colocation is not a cloud model at all and adds hardware work.

Correct answer: B — Private cloud, because the infrastructure serves one organization and still offers on-demand self-service.

A private cloud gives cloud behavior, including on-demand self-service provisioning, on infrastructure dedicated to a single organization, which is exactly what the regulator requires. Public cloud is ruled out because it is multi-tenant. Hybrid and multicloud describe where workloads are placed across environments; neither one on its own promises the dedicated infrastructure the regulator demands.

Correct answer: C — Agentic AI, because the system must plan toward a goal, call other systems as tools, and take several actions in sequence

Agentic AI is the right fit when a system must reason about a goal, choose its own next step, call tools or other systems, and act over several steps. That is exactly what this claims flow needs. Rule-based automation (A) breaks down here because the path changes for every claim and someone must code each rule. Generative AI alone (B) can draft text but does not take actions in other systems. A dashboard (D) reports on the past and automates nothing.

Correct answer: D — Cloud Run

Cloud Run runs a container image straight from a request, scales out automatically during bursts, scales to zero when idle, and leaves no cluster or node pool for the team to operate. A managed instance group still means machine images, health checks, and OS patching. GKE Standard adds cluster and node pool management on top of the application. GKE Autopilot removes node management but you still run and pay for a cluster and work with Kubernetes objects, so it is more overhead than Cloud Run for a single HTTP API.

Correct answer: C — Deploy only to Google Cloud regions in Germany, enforce the resource location constraint with an Organization Policy, and download the third-party audit reports from Compliance Reports Manager.

Residency is enforced by where resources may be created, and the Organization Policy resource location constraint blocks anyone from creating a resource outside the allowed German regions, which is the proof an auditor wants. Compliance Reports Manager then supplies the ISO, SOC, and other independent audit reports and certificates. Customer-managed keys give control over encryption keys, but they do not stop a bucket being created in another country, and a hand-kept spreadsheet is not third-party evidence. A multi-region location deliberately spreads data across a continent, which breaks a single-country requirement. VPC Service Controls reduces the risk of data leaving your perimeter, and the Transparency Report covers government requests for user data, so neither one is residency enforcement or an audit report.

Correct answer: D — Compute Engine

Compute Engine gives full control of the virtual machine: you pick the image, install a specific kernel and a vendor GPU driver, and keep a stable machine identity that a node-locked licence can bind to. Cloud Run and Cloud Run functions run your container or code on a fully managed platform where the host operating system and kernel are Google's, so a custom kernel or host driver is impossible. GKE Autopilot manages the nodes on your behalf, which is the point of Autopilot but also means you cannot pin a kernel version or install host-level drivers.

Correct answer: B — Hybrid cloud, by keeping the ERP on-premises and connecting the data center to Google Cloud

Hybrid cloud means part of the workload stays in a private data center while the rest runs in a public cloud, with the two connected. That matches a licensed ERP that must stay put plus analytics in BigQuery. Multicloud means using two or more public cloud providers, which is not what is happening here. Public cloud only fails because the license forbids moving the ERP, and private cloud only gives up BigQuery, which is the reason for the project.

Correct answer: C — Speech-to-Text API

The Speech-to-Text API converts recorded or streaming audio into text, which is exactly what makes the calls searchable. The Cloud Vision API analyzes images and video frames, so it has nothing to work with here. The Cloud Translation API moves text between languages, but the calls are already in English and are not text yet. Text-to-Speech is the reverse job: it turns written text into spoken audio.

Correct answer: A — Model Armor

Model Armor sits between the application and the model and inspects prompts and responses for prompt injection and jailbreak attempts, harmful content, and sensitive data, and its findings feed AI Protection in Security Command Center for overall AI posture. Cloud Armor is a network-edge WAF and DDoS defence that never sees the content of a model prompt. Identity-Aware Proxy controls who may reach the application, not what a legitimate user types into it. Sensitive Data Protection is the right tool for discovering and de-identifying sensitive data in storage, but scanning tables does not block an injected prompt or catch a leak in a generated reply.

Correct answer: D — 2-Step Verification is authentication, IAM roles are authorization, Cloud Audit Logs are auditing

Authentication proves who you are, and 2-Step Verification strengthens that proof by requiring a second factor at sign-in. Authorization decides what an identity may do, which is the job of IAM roles and permissions. Auditing records what was actually done, which is what Cloud Audit Logs capture for admin activity and data access. A, B, and C each swap at least two of these, and confusing authentication with authorization is the most common mistake on this topic.

Correct answer: C — Archive storage

Archive has the lowest price per GB and is built for data read less than once a year, which fits an eight-year compliance set touched annually. The trap is that Archive carries a 365-day minimum storage duration and the highest retrieval fee, so it only pays off when data is kept for years and read rarely, as here. Coldline targets data read about once a quarter and has a 90-day minimum, so it costs more at rest for no benefit. Nearline (30-day minimum) suits monthly access, and Standard is the most expensive place to park data that is almost never read.

Correct answer: A, D — Deploy the application across two or more zones in the same region so it keeps running if one data center fails · Use Cloud CDN so cached content is served from edge locations close to the users

A zone is roughly one data center inside a region, so spreading instances across zones is what protects you from losing a single data center. Edge locations sit in Google's global network close to users, so Cloud CDN cuts the distance a request travels and lowers latency. Live migration only moves a VM during host maintenance, it does not save you from a zone outage. Bigger machines do not shorten physical network distance, and forcing everyone to the head office region makes latency worse for users elsewhere.

Correct answer: C — Reading a customer support ticket, checking the order history, applying the refund policy, and updating the order system.

An agent earns its cost when the work starts from messy, unstructured input, needs several steps, and calls other systems along a path that is not known in advance. Handling a support ticket end to end fits all three: the agent must understand free text, decide what to look up, reason against policy, and act. Option A is reporting, which a dashboard already does better and more cheaply. Option B is a single threshold, which is a simple alerting rule. Option D is deterministic automation that a scheduled job handles reliably. Using an agent for those adds cost and unpredictability with no benefit.

Correct answer: A, E — The repeated customer rows are a uniqueness problem. · The age value of 400 is a validity problem.

Uniqueness means each real-world thing is recorded only once, so three rows for one customer is a uniqueness failure. Validity means values fall inside the allowed range or format, so an age of 400 is a validity failure. Completeness is about missing values, not duplicated ones, so B is wrong. The stale inventory feed is a timeliness problem, which is why C and D are traps: C calls a timeliness issue validity, and D calls a validity issue timeliness.

Correct answer: C — AI Hypercomputer, Google's integrated system of TPUs, GPUs, storage, networking, and software tuned to work together.

AI Hypercomputer is Google's supercomputing architecture for AI. Hardware, storage, networking, and the software stack are designed as one system, and that is the differentiator Google cites for price-performance on large-scale training and serving. The private global network is a genuine Google strength, but it delivers a consistent experience for users around the world, not cheaper training. Live migration improves availability during maintenance, not cost per training step. Committed use discounts apply to compute resources, not to Cloud Storage.

Correct answer: A, D — An SLI is the measurement itself, such as the percentage of requests that return successfully. · An SLA is the agreement with customers and usually carries a financial penalty when it is missed.

A service level indicator is the measured number, such as request success rate or latency. A service level agreement is the contract with customers and is the only one of the three that carries a financial consequence, such as service credits. B and C swap the two terms: the internal target is the SLO, and the contract with the penalty is the SLA. E is wrong because the error budget is the amount of unreliability the SLO allows, and that budget, not the SLA, is what tells the team when to stop shipping features and fix reliability.

Correct answer: D — Use AlloyDB for PostgreSQL in Google Cloud and run AlloyDB Omni in the factory.

AlloyDB for PostgreSQL is PostgreSQL compatible and adds a columnar engine that makes analytical queries on live transactional data far faster than standard PostgreSQL, which justifies its higher cost here. AlloyDB Omni is the downloadable edition of the same engine, so the factory runs the identical database on its own hardware. Cloud SQL for PostgreSQL is cheaper but does not give the analytical speed-up, and pairing it with community PostgreSQL means two different engines to manage. Spanner is not PostgreSQL and cannot run in the factory, and BigQuery is an analytics warehouse, not a transactional database.

Correct answer: D — A classic machine learning model trained on the delivery history to predict late parcels.

Predicting a specific outcome for a new record from labeled historical data is a classic machine learning task, so a trained prediction model is the right fit. A dashboard and a scheduled count both describe what already happened; they are business intelligence and cannot score a parcel booked today. Generative AI creates new content such as email text, which is useful afterwards but does not produce the prediction the managers asked for.

Correct answer: A — Each service can be released on its own schedule and the managed database removes routine patching and backup work, but the system is now distributed and is harder to test and troubleshoot end to end.

The real business value is speed and less toil: small teams ship their own service without a coordinated big-bang release, and a managed database takes patching, backups, and failover off the team's plate. The honest cost is complexity, because one user request now crosses many services and network calls, so testing and troubleshooting get harder. Option B is wrong because microservices usually add overhead and managed services still carry cost. Option C misreads shared responsibility: Google secures the infrastructure, but access control and your recovery requirements stay yours. Option D is the opposite of the truth, since distributed systems need more observability, not less.

Correct answer: C — Choose TPUs or GPUs to suit the model, then buy that capacity with Spot or committed use pricing.

AI Hypercomputer combines purpose-built accelerators, TPUs and GPUs, with flexible consumption models, so a workload that tolerates interruption can run on Spot capacity while steady baseline usage is covered by committed use discounts. Coldline storage cuts a small storage bill, not the accelerator cost that dominates training. Cloud Run is not built for large multi-accelerator training jobs, and a Vertex AI endpoint controls serving cost after training, not the training cost itself.

Correct answer: A, B — DNS translates a domain name into an IP address, so a slow or failing DNS lookup delays the very start of a request. · Bandwidth is the amount of data a network link can carry in a fixed amount of time, so a link with too little bandwidth slows large downloads.

DNS is name resolution: the browser must turn the domain name into an IP address before the first byte moves, so a slow resolver shows up as a slow app. Bandwidth is capacity per unit of time and explains slow bulk transfers. Option C describes concurrency, not latency; latency is the round-trip delay, and you reduce it by serving users from a region closer to them. Option D is wrong because DNS does no compression. Option E swaps the two terms: the millisecond delay is latency, not bandwidth.

Correct answer: C — Cloud Trace

Cloud Trace follows one request across services and records how long each span takes, so it answers exactly the question being asked: where in the call path the eight seconds go. Cloud Monitoring reports metrics and fires alerts, so it tells you the endpoint is slow but not which hop is at fault. Cloud Logging is for searching log entries and is most useful once you know which service to inspect. Cloud Profiler shows CPU and memory hot spots inside a single service, which is the right next step after Trace names the service. Error Reporting groups crashes and exceptions, and this request is slow rather than failing.

Correct answer: C — At the folder that contains the marketing department's projects

Policies flow down the resource hierarchy from organization to folder to project, so a policy set on the marketing folder is inherited by every project in it, including projects created later. The organization node would push the policy onto every other department too, which the requirement forbids. Setting it project by project works today but must be repeated each month, which is the effort the team wants to avoid. A Cloud Billing account handles payment and is not a parent in the resource hierarchy, so policies do not inherit from it.

Correct answer: B — In Cloud Storage buckets, used as a data lake for raw files in their original format

A data lake keeps raw, mixed-format data as-is at low cost so the schema can be decided later, and Cloud Storage is the usual landing zone for that on Google Cloud. Cloud SQL is a transactional relational database that needs a schema up front and does not suit high-volume raw logs. BigQuery is a data warehouse for analyzing structured data, and forcing a fixed schema before anyone knows the questions is the exact trap here; it is the next step after the lake, not the first. Memorystore is an in-memory cache, so it is not durable long-term storage.

Correct answer: A — The bank cannot justify an individual decision, which creates compliance and trust risk; responsible AI practices such as feature-level explanations and model documentation address it

When an automated decision affects a person, the organization has to be able to say which factors drove that specific outcome. Explainability tools such as Vertex Explainable AI produce feature attributions per prediction, and model cards record intended use and limits, so the bank can answer both the regulator and the customer. B is a different control: residency protects where data lives, not why a decision was made. C fails because aggregate accuracy says nothing about one applicant. D improves the model but still leaves the decision unexplained.

Correct answer: C — Train a Vertex AI AutoML image model on the company's own labeled photos

AutoML sits between a pre-trained API and a fully custom model: the company supplies its own labeled data and gets a differentiated model without deep learning skills or a long build. The Cloud Vision API is pre-trained on general objects, so it will never know this factory's defect categories. Custom training gives the most control and the most differentiation, but it needs specialists and far more time than the deadline allows. Document AI is built to read documents and forms, not to grade manufactured parts.

Correct answer: B — Encryption in use, met by Confidential Computing, which keeps data encrypted in memory while it is processed

Data has three states: at rest on disk, in transit over the network, and in use in memory. Confidential Computing, for example Confidential VMs, uses hardware-based trusted execution so memory stays encrypted while the workload runs, which is what the compliance officer asked for. A and D both protect stored data only, and the data would still be decrypted in memory during processing. C protects data moving over the network and says nothing about processing.

Correct answer: D — Spanner in a multi-region configuration

Spanner is the relational service built for horizontal write scale with strong consistency across regions, which is exactly what a growing payments ledger needs. Cloud SQL read replicas only spread reads; all writes still land on one primary instance, so the write ceiling stays. Bigtable is a wide-column NoSQL store without relational SQL joins or multi-row ACID transactions in the way a ledger expects. Firestore is a document database aimed at app data, not a relational ledger with SQL reporting.

Correct answer: A, E — Give each dataset a named owner who is accountable for its definitions and for running agreed quality checks. · Capture data lineage so that any figure in a report can be traced back to the source tables it came from.

Trust comes from knowing who owns a number and where it came from. Clear ownership fixes the two-teams-two-answers problem by making one team accountable for the revenue definition and its quality checks, and lineage lets anyone follow a figure back through the transformations to the source. Option B removes control instead of adding it and lets anyone change data with no review. Option C makes an untrusted number appear faster. Option D freezes the report format but does nothing about how the number was produced.

Correct answer: C — Elasticity: capacity is added during peaks and removed again when demand drops.

Elasticity is the ability to add resources automatically when demand rises and release them when demand falls, so the retailer stops paying for idle capacity between peaks. Scalability only describes the ability to grow, and it still leaves the off-season waste, so B fixes half the problem. Agility is about release speed, not capacity, and high availability is about surviving failures, not about matching demand.

Correct answer: B — Upfront purchases of servers and racks, plus the power, cooling, and three-year hardware refresh that go with them.

Owned hardware is the classic capital expense: you pay for capacity years ahead of the demand, then pay again for the power, cooling, floor space, and the refresh cycle. Cloud replaces that with pay-as-you-go operating expense, so you pay for what you use. Engineer salaries, third-party software licences, and staff training all remain after the migration, and training costs usually rise for a while during the move, so none of them is what changes from capital to operating expenditure.

Correct answer: A, B — Containers start in seconds because they share the host operating system kernel, while each virtual machine boots its own guest operating system · A container image bundles the application with its dependencies, so the same image runs on a developer laptop, on-premises, or on Google Kubernetes Engine

Containers virtualize at the operating system level and share the host kernel, so they start quickly and pack more workloads onto the same machine. The image also carries the app and its libraries, which is what makes it run the same way everywhere. C is the opposite of how containers work, since not carrying a guest OS is exactly why density is higher. D is wrong because VM image formats and drivers differ by platform and usually need conversion. E is wrong because containers still run on a host operating system.

Correct answer: A — Retire the unused expenses tool, retain the licensing application on-premises, and migrate the web application.

The point of discovery and assessment is to sort applications before spending migration effort: an application nobody uses is retired, an application with a hard blocker is retained where it is for now, and a clean, containerized application is moved. B wastes money migrating a dead application and cannot move the one with the hardware lock. C reverses the two decisions, retiring a working licensing system and keeping the unused tool. D forces expensive refactoring on every application, including one that should simply be switched off.

Correct answer: C — Publish the service through Apigee API Management and use its API keys, quotas, analytics, and monetization features

Apigee is Google Cloud's full API management platform. It issues and validates API keys or OAuth tokens per partner, enforces quotas and spike arrest, records per-partner analytics, and includes monetization for billing API consumers, so almost none of this has to be built. A means writing and maintaining all of that logic yourself. B can block or throttle traffic by rule, but Cloud Armor gives no developer credentials, per-partner usage reporting, or billing. D only caches responses and does not authenticate, meter, or monetize anything.

Correct answer: A — Lower quota limits for Compute Engine in the project, because quotas make the service reject requests beyond the limit while budget alerts only notify

Budgets and their threshold rules are a notification tool. They send an alert when spending crosses a percentage, but they never block a request or turn resources off, which is why the script kept running. Quotas are the enforcement control: when a project reaches its quota for a resource such as CPUs, further create requests are refused. B and C only change when and to whom the same notification is sent. D improves visibility after the fact and still stops nothing.

Correct answer: A — Bigtable for the sensor ingest, and Firestore for the mobile app

Bigtable is a wide-column NoSQL database built for very high write throughput and low-latency key lookups, which is the classic pattern for time-series device data. Firestore is a document database with client SDKs that cache data on the device and sync automatically when the network returns, so it fits the offline mobile app. B swaps them. C fails because a single relational instance cannot absorb hundreds of thousands of writes per second and gives no offline sync. In D, BigQuery is an analytics warehouse rather than a millisecond key-value store, and Cloud SQL offers no built-in offline mobile sync.

Correct answer: C — Sensitive Data Protection, to discover the personal data and de-identify it before analysts query it.

Sensitive Data Protection inspects text and structured data to find sensitive types such as names, phone numbers, and card numbers, and it can mask, tokenize, or redact them, which is what lets analysts keep the tickets without seeing the personal details. Cloud KMS encrypts data at rest, but an analyst with query access still reads the plain values. Cloud Armor filters inbound web traffic, not data already in BigQuery. VPC Service Controls stops data leaving a perimeter, but does nothing about approved analysts inside the perimeter reading the personal data.

Correct answer: A, B — Pub/Sub receives the events and decouples the store terminals from the consumers, buffering messages until they are processed · Dataflow reads the stream and applies the cleaning, windowing, and aggregation before writing results to BigQuery

Pub/Sub is the messaging layer: publishers write to a topic and subscribers read at their own pace, so a slow consumer never blocks the stores. Dataflow is the processing layer built on Apache Beam, and it applies transformations such as filtering, windowing, and aggregation before writing to a sink like BigQuery. C is the common trap, because Pub/Sub only delivers messages and does not transform or aggregate them. D reverses the two services. E adds a batch hop that is unnecessary and would break the near real-time requirement.

Correct answer: B — Define the metric once in a Looker semantic model and have all dashboards and reports read from that model

Looker holds business logic in a governed semantic model, so a metric such as active subscriber is defined once and every dashboard, report, and exploration reuses that definition. C does not solve the problem, because sharing the same tables still lets each team write its own filter. A relies on people keeping documents in step and does nothing to enforce the definition. D makes the problem worse by creating three drifting copies of the data.

Correct answer: B — Build the agent in Agent Studio on Agent Platform, ground it on the policy documents, and give it a tool that calls the claims system.

Agent Studio on Agent Platform is the guided environment for building agents on Google Cloud. Grounding an agent on your own documents and connecting it to internal systems through tools are built-in features, so a small team gets a working, governed agent quickly. Training a foundation model from scratch costs orders of magnitude more and needs a large ML team plus a huge dataset. Pasting documents into a public chatbot sends company data outside the company and cannot reach the claims system. Building on Compute Engine is technically possible, but the team then owns the retrieval pipeline, scaling, and security, which is the slowest option here.

Correct answer: A — Place the application behind Identity-Aware Proxy and grant access to each contractor's identity.

Identity-Aware Proxy checks identity and context on every request to the application itself, so contractors reach only that one application and never join the network, and no VPN client is needed. A VPN puts each contractor inside the network, which grants far more reach than the single application requires. Filtering by office IP address trusts a network location rather than a person and breaks when contractors work from home. Sharing service account keys is a serious anti-pattern: the credential is long-lived, shared, and cannot identify who used it.

Correct answer: B — Autoscaling fixes the peak-hour timeouts; load balancing fixes the uneven load across machines.

Autoscaling changes how many instances exist, so it is what removes peak-hour timeouts caused by too little capacity. A load balancer distributes incoming requests across the instances that already exist, so it is what stops one instance running hot while others idle. C swaps the two roles. A and D each claim one service does both jobs, but a load balancer never creates instances and autoscaling never decides which instance receives a given request.

Correct answer: A — Create a BigQuery ML time series model with SQL and run the forecast inside BigQuery

BigQuery ML lets analysts create, train, and serve a forecasting model with CREATE MODEL and ML.FORECAST statements, so the data never leaves BigQuery and no new platform or language is needed. Custom TensorFlow training gives the most control but demands ML engineers the company does not have. An AutoML forecasting model would work, yet it means setting up a separate platform and workflow when SQL alone is enough. A Looker Studio trend line is only a chart decoration; it is not a trained forecasting model and gives no confidence intervals.

Correct answer: B — Run the rendering job on Spot VMs and keep the database on standard VMs, because Spot VMs can be reclaimed at short notice

Spot VMs cost far less than on-demand instances but Compute Engine can preempt and shut them down at any time with only a short warning, and it may not have capacity when you ask for them. That failure mode is fine for a fault-tolerant, restartable batch render and unacceptable for a stateful production database that must stay up. A puts the portal at risk of being preempted mid-transaction. C is the reverse of the correct mapping. D leaves easy savings on the table and states a rule that does not exist.

Correct answer: A, C — Under IaaS, the customer patches the guest operating system on its Compute Engine virtual machines · Under every service model, the customer classifies its own data and decides who may access it

With IaaS, the provider secures the hardware and hypervisor, but everything from the guest OS upward, including patching, belongs to the customer. Data classification and access decisions never transfer: only the customer knows how sensitive its data is and who should see it, so this holds under IaaS, PaaS, and SaaS. Under PaaS the provider patches the underlying OS and runtime, which is the whole point of the model. Under SaaS the provider owns the application code, and under IaaS the customer, not Google, owns the application it deploys on the VM.

Correct answer: D — Gemini Enterprise, which gives employees a governed platform to find, build, and run AI agents connected to company data sources

Gemini Enterprise is Google Cloud's enterprise agent platform. It connects to company data sources, lets business users build and share agents with low-code tooling, and gives IT central governance over access. A is only the content store and does not provide agent building or connections to the ticketing system and databases. B accelerates BI queries and has nothing to do with agents. C would work only after the company builds and maintains the connectors, security, and governance itself, which is far more operational effort.

Correct answer: B — Open source is code anyone can inspect, run, and modify; an open standard is a public specification that many vendors implement, so work built to it moves between providers

Open source is about the code: it is published under a license that lets you inspect, run, and change it, so you can host it yourself or elsewhere. An open standard is about the specification, such as SQL or the Kubernetes API, and many vendors implement it, so skills and workloads carry across providers. A is wrong because the two terms are not the same, and open source is not defined by price. C reverses the definitions. D is false: open source software can be run on any provider or on your own hardware.

Correct answer: B — Attach a Google Cloud Armor security policy with preconfigured WAF rules to the load balancer

Google Cloud Armor works with the external Application Load Balancer, so it can inspect HTTP requests, absorb layer 7 volumetric attacks at the edge, and block SQL injection with preconfigured WAF rules. VPC firewall rules only match IP address, protocol, and port; they cannot read a query string, and chasing thousands of changing source IPs does not scale. Cloud NAT controls outbound access and does nothing about inbound attacks. Hierarchical firewall policies are still layer 3 and 4 rules, just applied higher in the resource hierarchy.

Correct answer: C — Use BigQuery Omni to query the data where it already sits in Amazon S3.

BigQuery Omni runs BigQuery compute in AWS and Azure, so analysts write normal BigQuery SQL while the raw data stays in place and only query results cross the boundary. Storage Transfer Service, the BigQuery Data Transfer Service, and a Dataflow pipeline all copy the claims data into Google Cloud, which is precisely what the residency rule forbids.

Correct answer: B — Adopt Google Security Operations as a single platform for detection, investigation, and response, enriched by Google Threat Intelligence.

Google Security Operations brings SIEM and SOAR together, so telemetry from cloud, on-premises, and third-party tools lands in one place and analysts stop pivoting between consoles. Google Threat Intelligence supplies what makes those detections current: Mandiant frontline incident response, VirusTotal's global malware submissions, and Google's own visibility across billions of users and devices. Cloud Armor defends the network edge and does nothing for investigation workflow. Logs parked in a bucket are evidence, not detection, and querying them after the fact is exactly the slow path the bank wants to leave. Security Command Center is valuable for cloud posture and misconfiguration risk, but it does not cover on-premises and third-party telemetry or replace a SOC platform.