PCA — frequently asked questions

BigQuery is the analytics destination, but the scenario describes processing requirements — hourly jobs plus live-processing data as it arrives. Dataflow performs that unified batch and stream processing, after which results can be loaded into BigQuery for analysis.

Spark on Dataproc can run both, but Dataproc requires you to provision and manage clusters. With no existing code, Dataflow's serverless Beam model is the simpler, officially preferred answer.

Object storage has no rows or query engine and is built for batch-oriented file access, not a continuous per-click ingest stream. Bigtable handles the event-rate write volume and still supports the later analytical reads.

No — Cloud SQL is a relational OLTP service whose write throughput, connection count, and storage growth scale vertically and saturate far below a sustained event stream of this size.

A canary sends only a small share of traffic, so load-dependent performance bugs may never appear until full rollout. It limits blast radius but does not reproduce production load during testing.

The question permits adjusting either test or deployment procedures; fixing unrealistic test load removes the root cause before any deployment occurs.

While GCS is cheaper, it stores logs as flat files. Analyzing them requires custom scripts or manual inspection, which is slower and less scalable than using BigQuery's SQL interface for instant filtering.

Yes, there are storage and query costs associated with BigQuery. However, for organizations requiring frequent or detailed audits, the operational efficiency and reduced labor time often justify the expense.

Deployment Manager provisions Google Cloud infrastructure from templates (VMs, networks, disks); it does not create Kubernetes workload objects. A Deployment manifest must be submitted through kubectl against a running cluster's API server.

kubectl is a client for an existing Kubernetes API endpoint, so it cannot provision a GKE cluster. Creating a cluster on Google Cloud is done with gcloud container clusters create, the Console, or Terraform.

Google Cloud documentation states that read replicas do not provide failover capability; they are for read scaling. For HA failover, you need a regional HA configuration with a standby instance in another zone.

Yes. Simulating a zone failure directly tests whether the regional MIG can withstand the loss of compute resources in one zone, which is a core principle of resilience design.

While useful for monitoring, INFORMATION_SCHEMA tables are not retained as long as audit logs and are not considered an immutable audit trail for compliance.

No, bq ls lists jobs interactively and is not practical for aggregating historical data across multiple users and months programmatically.

BigQuery is an analytics warehouse built for batch loads and SQL scans; it cannot sustain continuous high-rate ingestion with millisecond point lookups, which Cloud Bigtable delivers.

Traffic has idle periods and cost must stay low, so Cloud Run's ability to scale to zero matters; a managed instance group must keep at least one VM running and billing.

Cloud CDN caches content at Google's edge locations and integrates directly with the HTTP(S) Load Balancer the question requires, delivering lower latency and higher reliability than serving straight from a bucket.

Cloud Run is purpose-built for containerized HTTP services with longer request timeouts and streaming, whereas Cloud Functions targets short-lived event-driven workloads and isn't ideal for public API backends.

At 1 Gbps, 10 TB transfers online in about 30 hours. Transfer Appliance requires shipping, loading, returning, and rehydrating data—totaling 1–3 weeks—so it is substantially slower for this volume and bandwidth.

That recommendation is conditional: appliance is preferred only when the online transfer time substantially exceeds the ~20-day appliance turnaround. Here 30 hours is far less than 20 days, so gsutil -m is the faster, cheaper option.

No, snapshots are project-scoped. You must either share the snapshot via IAM permissions or convert it to an Image and import it into Project B.

Option B implies using the snapshot directly. In GCP, you cannot attach a snapshot from Project A as a boot disk for an instance in Project B without first creating an Image from that snapshot.

Writing through gcsfuse is network-bound and far slower than Local SSD I/O, so the backup runs longer and the production database stays under read load for more time. Local SSD staging plus gsutil gives the fastest completion, which the question explicitly demands.

Cloud Persistent Disk snapshots capture an entire disk image, not a single logical database, and the snapshot process reads from the same persistent disk serving production traffic. That violates the requirement that backup activity must not impact disk performance.

Bigtable's performance depends on your schema design, row key layout, cluster sizing, and access patterns. Google's own documentation recommends running your typical workloads against Bigtable for capacity planning, so you must validate it under load.

Running load tests against production risks degrading real users, and recording every production transaction for replay is impractical and raises privacy concerns. Load testing should use a dedicated isolated environment with synthetic or anonymized data.

Appliances are recommended only if online transfer exceeds one week. On a 1 Gbps link, 10TB takes ~30 hours, making the appliance slower and more expensive.

No. The 5TB limit applies to single objects. You can split the database dump into multiple smaller files and upload them all concurrently using gsutil/gcloud.

Monitoring alerts track performance and uptime. While budget alerts exist, they lack the granular visibility needed to understand which resources drive costs without proper labeling.

While technically possible with custom pipelines, GCP's native and recommended workflow for detailed billing analysis is exporting to BigQuery, making option C incorrect in this context.

Archive restores take minutes to hours and its retrieval fee is far higher than Coldline's, so quarterly reports would miss the acceptable few-second retrieval window. The storage saving is offset by retrieval cost at terabyte scale.

No. The minimum duration only penalizes objects deleted or reclassified before 90 days, and this data is retained for five years, so the charge never applies.

VPA Auto mode applies new recommendations by evicting and restarting Pods, which can disrupt stateful Dealer Tools workloads and cause downtime. It also reacts to sustained usage trends rather than sudden load spikes.

Requests equal to limits creates a Guaranteed Pod that reserves its full allocation permanently, keeping the over-allocation problem intact and reducing how many Pods fit on each node.

Replicas provide near-continuous data synchronization, keeping data loss within minutes (15 min RPO). Snapshots are point-in-time and can lose up to a day's worth of data if taken daily.

Yes, Google Cloud supports Global External TCP/UDP Load Balancing. The question specifies HTTP(S), but the principle of global health-check-based failover applies similarly to other protocols.

Option A protects against a full regional outage, which exceeds the stated requirement, and its global load balancer plus multi-regional Spanner costs far more than needed. The minimize-cost constraint rules it out.

No. Spot VMs can be preempted with little notice, so they cannot guarantee availability for a critical payment workload, even though the regional MIG would rebalance capacity.

Dataproc is managed Hadoop/Spark, but you still size, autoscale, patch, and monitor the clusters, which adds the operational overhead the question asks to minimize. Dataflow provides the same unified stream-and-batch execution (via Apache Beam) with no infrastructure to manage.

No. Dataprep is a data preparation tool, not a processing engine, so it cannot run real-time fraud detection or batch analytics. Firestore is a NoSQL document store for low-latency app serving; it cannot support large-scale analytical scans and SQL aggregation on transaction data.

They are complementary. Dataflow processes and transforms the streaming (and batch) data, while BigQuery stores and queries the data for the complex daily reporting; BigQuery also offers streaming ingestion, but the processing/orchestration role belongs to Dataflow here.

Concurrency increases how many requests one instance handles at once, improving throughput under load, but it does not move images, videos, or data closer to the customer, so page load speed is unchanged.

Changing databases does not reduce the number of network hops a customer request makes; caching at the edge and in memory is what actually cuts the time to first byte for media-heavy pages.

Cloud SQL high availability is regional only; true multi-region resilience needs cross-region read replicas plus manual failover, which breaks both the HA and low-operational-overhead requirements.

Firestore is a document database and Bigtable is a wide-column NoSQL store; neither provides full relational semantics such as joins, foreign keys, and standard SQL transactions.

Cloud Run is serverless containers — no cluster, node pool, or control-plane management — so it minimizes operational overhead while a global HTTPS load balancer still provides multi-region entry.

Colab Enterprise is a valid managed Jupyter option, but pairing it with generic 'DevOps' omits a Google-recommended MLOps deployment mechanism, so the option fails the deployment half of the requirement.

AutoML automates model building and limits custom training loops and fine-tuning, which contradicts the stated need for maximum flexibility for ML engineers doing custom data preparation and training.

Google explicitly prohibits committing service account keys to version control because anyone with repo access could reuse the credential. Cloud Build's GitHub App authenticates via IAM instead, eliminating long-lived keys.

Config Sync is a GitOps tool that reconciles Kubernetes cluster configuration from a Git source; Cloud Run is a serverless container platform, not a Config Sync-managed cluster. Cloud Build should deploy to Cloud Run directly.

Search results may be outdated, out of scope, or unofficial, so they cannot serve as reliable audit evidence. The Compliance Reports Manager delivers verified reports tied to your Google Cloud account.

Cloud Audit Logs captures activity inside your own projects, such as Admin Activity and Data Access entries. It does not distribute external certification reports or attestations.

Zonal instances are independent and do not replicate data, so the VMs would see several separate file shares instead of one shared file system, forcing manual partitioning or syncing of the raw files.

Cloud Storage is object storage with no mounted POSIX file share and no low-latency concurrent read/write semantics for hundreds of VMs, so it cannot serve as the shared file system the pipeline requires.

Spanner is for relational data needing strong consistency and complex joins. Datastore is more cost-effective and flexible for simple key-value profile storage.

Yes, but it is often referred to as Firestore in Datastore mode. For exam purposes, 'Datastore' remains the standard term for this NoSQL capability.

While Multi-Regional offers higher availability, it has lower write throughput limits and higher write latency due to cross-region replication. For ETL ingestion, speed and reliability of the initial upload are prioritized, making Regional buckets the better fit.

No. Directly uploading to Cloud Storage via HTTPS APIs is more secure, scalable, and reliable than hosting custom FTP servers on VMs, which adds management overhead and potential single points of failure.

Because moving the full uncompressed raw dataset intercontinentally costs about $0.12 per GB, which dominates the total bill. Compressing locally first shrinks the transfer volume dramatically before any cross-region copy.

Multi-region storage costs roughly 30% more per GB-month than regional storage (about USD 260 vs USD 200 for 10 TB/month), and Dataproc cannot run inside a multi-region anyway—the cluster must live in a specific region.

Cloud Dataproc clusters are regional resources; you choose a region (and optional zone placement for nodes) when you create the cluster, so a regional GCS bucket is the natural co-located storage target.

HA provides a synchronous standby in the same region to survive zone failures. Cross-region replicas provide an asynchronous copy in another region to survive regional outages.

Automated backups have too high an RTO and RPO for strict requirements like <10 min RTO and <2 min RPO. They are better suited for long-term retention or less critical workloads.

Local SSDs offer lower latency but are ephemeral. When a Spot VM is preempted, all data on local SSDs is lost. For batch jobs requiring data persistence, Persistent Disks are necessary, often paired with checkpoints.

Preemptible VMs have been deprecated and replaced by Spot VMs. Spot VMs offer similar interruption-tolerant, low-cost benefits but with more predictable availability and better integration with modern Google Cloud features.

1,000 writes per second is only about 86 million rows per day, which managed relational databases handle easily. The one-second cadence is not true low-latency pressure, and the payload is described as discrete, structured fields rather than unknown-shape data.

That join is the actual analyst workload: sensor readings are facts that must be related to owner and office dimensions. A relational engine executes that join in the query engine with indexes, while NoSQL would require denormalizing those fields into every reading.

Both store bytes but provide no indexing, querying or join capability, so analysts could not run the required reporting against them. They are storage tiers, not analytical databases.

Because the script already runs and reaches the connection step — a missing library would fail at import time with ModuleNotFoundError. The failure is authorization, not missing dependencies.

Access scopes are a legacy mechanism set at instance creation and only define default OAuth scopes for the VM's requests. Google recommends granting IAM roles to the instance's service account instead.

Attach the service account to the instance and use Application Default Credentials, or export GOOGLE_APPLICATION_CREDENTIALS pointing to a downloaded service account key file before running the script.

Use Transfer Appliance when the data set is at least 10 TB (20 TB in newer docs) and uploading over your link would take more than a week. Smaller sets, typically a few TB, are fine with gsutil resumable uploads.

It is the Google-side tool that takes the encrypted contents received on the appliance and decrypts them directly into your destination Cloud Storage bucket, completing the offline transfer workflow.

Cloud Dataprep profiles and prepares tabular data; it neither ingests bulk data sets nor decrypts a Transfer Appliance, so it cannot move 75 TB into Cloud Storage.

Yes, Cloud SQL HA keeps a standby in another zone of the same region and fails over automatically when the primary zone fails. However, the active-passive switchover pauses writes for the duration of the failover, so it reduces downtime rather than eliminating it.

Cloud Spanner is Google's own horizontally scalable SQL database and is not Microsoft SQL Server. It fails the explicit "set up Microsoft SQL Server" requirement even though a regional Spanner instance is highly available.

In GCP, subnets are regional resources that can span every zone, so nodes in separate subnets may still share one zone and fail together. Placing replicas in different zones is what guarantees survival of a single data-center failure.

A multi-region bucket replicates objects across regions within the chosen multi-region, which improves availability and serving posture compared with a single region. For edge caching of public downloads, add Cloud CDN; the bucket location itself is the exam focus.

Cloud Storage buckets are not zonal resources. You choose a region, dual-region, or multi-region, so options B and C are based on an invalid placement model and create unnecessary copies.

Multiple multi-region buckets can be part of a custom global distribution design, but you must replicate content and route users to the right bucket. This question asks for Google-recommended storage without that routing layer, so A is the defensible answer.