Task 1.3: Designing network, storage, and compute resources.

• APeer the company's VPC with the vendor's VPC by using VPC Network Peering

  Incorrect

  Peering fails when subnet ranges overlap, and it exchanges routes for whole networks rather than one service.

• BAttach both projects as service projects of one Shared VPC host project

  Incorrect

  Shared VPC connects projects within one organization; it cannot join the vendor's projects in another organization.

• CBuild HA VPN tunnels between the two VPC networks and advertise the service subnet

  Incorrect

  VPN routing also needs non-overlapping ranges, and it connects networks rather than exposing a single service.

• DCreate a Private Service Connect endpoint for the vendor's service attachment

  Correct

  Private Service Connect gives the consumer an internal IP address in its own VPC for one published service, so overlapping ranges and the rest of the producer network do not matter.

Private Service Connect is the producer-consumer model for private access to a managed service. The consumer reaches the service through its own address space, traffic stays on Google's network, and nothing else in the producer's VPC becomes reachable.

• ABuild a time-series forecasting model with BigQuery ML and query it in SQL

  Correct

  BigQuery ML builds and runs models, including ARIMA_PLUS time-series forecasts, with GoogleSQL on data that stays in BigQuery.

• BExport the tables to Cloud Storage and train on Managed Service for Apache Spark

  Incorrect

  This moves the data out of the warehouse and needs Spark programming skills the analysts do not have.

• CReplicate the tables to Cloud SQL and forecast with scheduled stored procedures

  Incorrect

  Cloud SQL is a transactional database; it is not built for this analysis and the data would have to be copied.

• DTrain a custom forecasting model in Python notebooks in Agent Platform Workbench

  Incorrect

  Workbench notebooks suit Python data scientists, not analysts who work in SQL.

Choosing a data processing solution starts from where the data lives and the team's skills. BigQuery ML lets SQL practitioners create, evaluate and use forecasting models inside BigQuery.

• AA Cloud Storage bucket mounted with Cloud Storage FUSE

  Incorrect

  Cloud Storage FUSE gives a file interface over objects, but it is not POSIX compliant and lacks file locking.

• BA Persistent Disk attached to every VM in read-only mode

  Incorrect

  In read-only mode no VM can write, and the application writes to the shared tree.

• CA Filestore instance mounted over NFS on every VM

  Correct

  Filestore is managed NFS file storage with POSIX compliance, file locking and concurrent writes from many clients.

• DA Local SSD on each VM kept in step with rsync jobs

  Incorrect

  Local SSDs are per-VM scratch disks; copying between them gives no shared view and no locking.

Choose the storage type by access pattern. Many VMs sharing a POSIX file system with locking is a file workload, which on Google Cloud means managed file storage such as Filestore.

• ASupervised fine-tuning of a Gemini model on the PDF text every night

  Incorrect

  Nightly tuning is custom ML work, and a tuned model does not retrieve and cite the passages behind each answer.

• BA custom model pretrained on the PDFs with Cloud TPU capacity

  Incorrect

  Pretraining a model is the most custom ML work possible and is unnecessary for question answering over documents.

• CAn Agent Search data store that ingests and indexes the PDFs from Cloud Storage

  Correct

  Agent Search manages ingestion, parsing, indexing and retrieval of enterprise documents with no model training.

• DA Gemini model on Agent Platform grounded with Grounding with Agent Search

  Correct

  Grounding connects Gemini to the Agent Search data store so answers are based on the retrieved documents and can cite them.

• EA BigQuery ML k-means model that clusters the PDFs by topic

  Incorrect

  Clustering groups documents; it does not answer questions or cite passages.

Retrieval-augmented generation is the low-effort pattern for answering from private documents: a managed search index retrieves the relevant passages and a grounded model writes the answer from them, with citations.

• AA VPC per application project, connected to each other with VPC Network Peering

  Incorrect

  Each team would own its own network, so the network team would not control subnets, routes and policies centrally.

• BOne project that holds the network and every application team's resources together

  Incorrect

  A single project removes the isolation and separate IAM boundaries that the teams' own projects provide.

• CShared VPC, with a host project for the network team and service projects for apps

  Correct

  Shared VPC lets resources in service projects use subnets of a host project's network, so network administration stays central while teams keep their own projects.

• DPrivate Service Connect endpoints from each application VPC into one central VPC

  Incorrect

  Private Service Connect exposes individual services; it does not let teams place VMs in centrally managed subnets.

Shared VPC is the standard pattern for central network control with decentralized application ownership inside one organization.

• AMore VPC Network Peering connections until every pair of networks is peered

  Incorrect

  A full mesh of 30 networks means hundreds of peerings and runs into peering limits.

• BNetwork Connectivity Center, with each VPC network attached to one hub as a spoke

  Correct

  Network Connectivity Center connects VPC and hybrid spokes through a central hub, which gives full connectivity without a mesh of pairwise peerings.

• CCloud NAT gateways in every VPC network to carry the traffic between the networks

  Incorrect

  Cloud NAT translates addresses; it does not route traffic between VPC networks or to on-premises sites.

• DPrivate Google Access on every subnet in each of the 30 VPC networks

  Incorrect

  Private Google Access does not connect a subnet to other VPC networks or to on-premises sites.

Hub-and-spoke connectivity with Network Connectivity Center scales where pairwise peering does not, and it can include hybrid connections as spokes.

• ACloud SQL for MySQL with one readings table per vehicle model

  Incorrect

  A single relational instance cannot absorb millions of writes every few seconds.

• BBigQuery tables with streaming inserts as the lookup store

  Incorrect

  BigQuery is built for analytical scans, not millisecond lookups of one vehicle's rows.

• CBigtable, as a wide-column store for the readings of every vehicle

  Correct

  Bigtable handles very high write throughput with low-latency key lookups, which fits frequent small readings that are read back by vehicle.

• DMemorystore for Redis as the system of record for every reading

  Incorrect

  An in-memory cache is costly at this volume and is not meant to be the durable record.

Time-series data at massive write rates with key-based reads is the classic Bigtable workload.

• AGPUs on Compute Engine with custom CUDA kernels

  Incorrect

  GPUs suit models that rely on custom operations; this model has none, and its profile matches the TPU guidance.

• BCloud TPU slices sized for the model's training job

  Correct

  TPUs are built for large matrix operations and suit large models that train for weeks with no custom operations in the main loop.

• CCloud Run services configured with NVIDIA L4 GPUs

  Incorrect

  Cloud Run GPUs target inference and similar services, not weeks-long distributed training.

• DAutoML training jobs on Gemini Enterprise Agent Platform

  Incorrect

  AutoML trains models from labeled data without code; it does not train a custom JAX transformer.

Choosing compute for specialized workloads means matching the model's profile to the accelerator. Google's guidance points large, matrix-heavy models without custom operations to TPUs.