Free Google Professional Cloud Architect Sample Questions

Correct answer: B — Custom Machine Type.

• A. Incorrect. An N2-standard-32 provides 128 GB of memory but also 32 vCPUs. Since the application only requires 2 vCPUs, this results in significant over-provisioning and higher costs for unused compute resources.
• B. Correct. A Custom Machine Type on Compute Engine allows you to provision a specific number of vCPUs and memory. By using custom machine types with 'Extended Memory,' you can configure an instance with 2 vCPUs and 128 GB of RAM, optimizing costs by only paying for the resources needed while still allowing kernel-level modifications on a full VM.
• C. Incorrect. Sole-tenant nodes provide dedicated physical hardware for compliance, security, or licensing needs. While they support custom VM configurations, they are generally more expensive than shared-core or standard instances and are unnecessary for simply solving a CPU-to-memory ratio mismatch.
• D. Incorrect. Cloud Run is a serverless container platform. It does not support kernel-level modifications or specific legacy Linux OS versions, and the application requirements explicitly state it cannot be containerized.

Correct answer: C — Filestore High Scale or Enterprise.

• A. Cloud Storage FUSE allows Google Cloud Storage (object storage) to be mounted as a file system, but it does not provide full POSIX semantics and has significantly higher latency and lower throughput than required for HPC workloads. It is not optimized for simultaneous high-performance access by hundreds of compute nodes.
• B. Persistent Disk (pd-extreme) is high-performance block storage designed for individual instances. While it offers high IOPS, it is not a shared file system. Using it for shared access would require manual configuration of a distributed file system (like Lustre or GlusterFS) on top of VMs, which violates the requirement for a managed service to avoid patching servers.
• C. Filestore High Scale and Enterprise tiers are managed NFS services designed for performance-critical workloads. They provide the necessary throughput (multiple GB/s), low latency, and POSIX compliance required by HPC applications, and they natively support concurrent access from hundreds of compute nodes without the need for manual server management.
• D. The Google Drive API is designed for end-user document collaboration and individual file storage. It does not offer a POSIX-compliant interface, the high-throughput performance characteristics, or the low-latency connectivity required for an HPC simulation cluster.

Correct answer: B — Configure Object Lifecycle Management policies to transition objects to Nearline, Coldline, and Archive classes based on age.

• A. Writing a shell script is a manual, brittle approach that requires ongoing maintenance and scheduling. It does not scale effectively for petabytes of data and fails to leverage built-in cloud-native automation like Object Lifecycle Management.
• B. Configuring Object Lifecycle Management (OLM) policies is the recommended and most efficient way to automate cost optimization in Google Cloud Storage. OLM allows for the automatic transition of objects to colder storage classes (Nearline, Coldline, Archive) based on their age, ensuring the data is stored in the most cost-effective manner without manual intervention or custom code.
• C. The Transfer Appliance is designed for large-scale migrations into the cloud, not for ongoing cost optimization within the cloud. Moving data to an on-premises data center adds significant operational complexity, loses the benefits of cloud scalability, and does not provide an automated lifecycle solution.
• D. Attempting to compress petabytes of video footage using Cloud Functions is impractical due to resource limits (CPU/Memory) and execution time constraints. The compute cost of re-encoding or compressing large volumes of video would likely outweigh storage savings, and it does not address the need for automated storage tiering.

Correct answer: C — Use Terraform to provision resources and a configuration management tool (like Ansible or Chef) or Google Cloud Runtime Configurator to manage state.

• A. Startup scripts are imperative and ad-hoc rather than declarative. They run at boot time but do not provide state reconciliation or active management to prevent configuration drift once the instance is running. They lack the auditing, versioning, and idempotency guarantees required for a true IaC solution.
• B. Manual image creation is a labor-intensive, error-prone process that contradicts the goal of infrastructure as code. It reintroduces human intervention, does not scale effectively, and fails to provide a documented, version-controlled path for configuration changes.
• C. This is the most reliable approach for managing Compute Engine fleets. Terraform provides declarative infrastructure provisioning, while configuration management tools like Ansible or Chef provide idempotent software configuration (IIS). This combination ensures that the environment's state is defined in code, can be automatically reconciled to prevent drift, and supports version-controlled, repeatable deployments.
• D. While GKE with Windows containers provides excellent declarative management and immutability, it requires containerizing the applications. This represents a significant architectural shift and migration effort rather than an evolution of the existing Windows Server fleet management. For the specific goal of managing Compute Engine instances via IaC, option C is the standard architectural progression.

Correct answers: A, B — Use Compute Engine instances within a Managed Instance Group (MIG).; Configure the MIG to use Spot VMs.

• A. Managed Instance Groups (MIGs) provide self-healing capabilities. If an instance is deleted or reclaimed, the MIG automatically recreates it based on the instance template, ensuring that the stateless workload can be restarted on a new instance without manual intervention.
• B. Spot VMs are the most economical compute option available in Google Cloud, offering significant discounts compared to standard pricing. They are ideal for fault-tolerant and stateless workloads like batch rendering, where the application can handle the instance being reclaimed by Compute Engine.
• C. Sole-tenant nodes provide dedicated physical hardware for your instances. This is used for specific compliance, security, or licensing requirements and is more expensive than standard or Spot VMs, making it inappropriate for cost-minimization goals.
• D. Shielded VMs provide verifiable integrity through security features like Secure Boot and Virtual Trusted Platform Module (vTPM). While they enhance security, they do not impact cost optimization or the ability to restart reclaimed instances.
• E. Committed Use Discounts (CUDs) provide savings for long-term, steady-state usage (1 or 3 years). However, for fault-tolerant batch workloads, Spot VMs offer a much higher discount and are more appropriate than committing to standard VMs with a CUD.

Correct answer: D — Four Dedicated Interconnect connections in two different metropolitan areas (two in each metro).

• A. A single Dedicated Interconnect connection represents a single point of failure. Google does not offer an availability SLA for a single connection as it lacks any redundancy.
• B. Two Dedicated Interconnect connections in a single metropolitan area (deployed across two different edge availability domains) are the requirement for a 99.9% availability SLA, not the 99.99% required by the business.
• C. Two connections in two different metros is insufficient for the 99.99% SLA topology. While it provides geographic redundancy, Google Cloud requires redundant connections within each metro to protect against edge availability domain failures while maintaining connectivity to both metros.
• D. To achieve a 99.99% availability SLA, Google requires a specific topology: four Dedicated Interconnect connections. These must be deployed as two connections in one metropolitan area and two connections in a second metropolitan area (each pair using different edge availability domains). Additionally, Global Dynamic Routing must be enabled in the VPC.

Correct answer: C — Google Cloud Bare Metal Solution (BMS).

• A. Incorrect. Google Cloud SQL is a managed service that does not support Oracle Database; it supports engines like PostgreSQL, MySQL, and SQL Server. Furthermore, it is a multi-tenant service, which violates the requirement for dedicated infrastructure.
• B. Incorrect. While Sole-Tenant Nodes provide physical isolation from other customers, they still run virtualized instances on a hypervisor. Oracle's strict licensing policies often require non-virtualized hardware to maintain specific core counts efficiently and avoid licensing the entire physical host at a high cost.
• C. Correct. Google Cloud Bare Metal Solution (BMS) provides dedicated, non-virtualized physical servers. This is specifically designed for specialized workloads like Oracle that require direct access to hardware for performance and strict licensing compliance based on physical core counts, allowing the company to mirror their on-premises core configuration.
• D. Incorrect. Standard Compute Engine instances are part of a shared, multi-tenant infrastructure and use virtual CPUs (vCPUs) rather than guaranteed physical cores. This does not meet the requirements for license compliance or non-shared infrastructure.

Correct answer: B — Implement the Strangler Fig pattern: Build new features as microservices on Google Cloud and gradually route traffic away from the mainframe until it can be decommissioned.

• A. Rehosting using a mainframe emulator on Compute Engine is a lift-and-shift approach. While it may reduce data center hardware costs, it preserves the monolithic legacy architecture and its operational constraints, failing to achieve the stated goal of modernization to microservices.
• B. The Strangler Fig pattern is the industry-standard approach for migrating complex legacy systems like mainframes to microservices. It allows for the incremental replacement of functionality with new services on Google Cloud, reducing risk by routing traffic gradually and enabling continuous feature development while the mainframe is eventually decommissioned.
• C. This represents a 'big-bang' migration and rewrite strategy. Such an approach is extremely high-risk, time-consuming, and does not inherently lead to a microservices architecture. It focuses on database and language modernization but lacks the incremental decomposition required for complex system transformations.
• D. Google Cloud VMware Engine (GCVE) is designed to run VMware-based x86 workloads. Mainframe operating systems are not compatible with this environment, and this approach would not facilitate the transition to a modern microservices-based architecture.

Correct answer: C — Retire the legacy CRM and Buy (SaaS) a modern CRM solution (e.g., Salesforce) integrated via connectors.

• A. Rehosting (Lift and Shift) to Compute Engine moves the workload to the cloud but preserves the existing application architecture, licensing model, and operational burdens such as patching and manual upgrades. It yields limited long-term ROI because it fails to address the expensive licensing or the lack of mobile support.
• B. Replatforming to containers on GKE can improve portability and scalability, but it still requires maintaining the legacy application code and handling proprietary licensing. It typically requires substantial engineering effort and does not address the application's inherent lack of mobile support or modern features.
• C. Buying a modern SaaS CRM solution addresses all the stated pain points. It eliminates expensive legacy licensing and patching toil (offloading them to the vendor), provides built-in mobile support, and modernizes the user experience. For common business functions like CRM, SaaS typically yields the highest long-term ROI by reducing total cost of ownership and accelerating productivity.
• D. Refactoring a legacy proprietary CRM into microservices using Cloud Functions involves a massive upfront cost, development effort, and risk. While it maximizes cloud-native benefits, the time-to-value is much longer than a SaaS migration, and for a proprietary system, it may not be feasible or cost-effective compared to buying a mature market solution.

Correct answer: C — Regional External Passthrough Network Load Balancer

• A. Incorrect. The Global External Application Load Balancer is a Layer 7 load balancer designed specifically for HTTP(S) traffic. It does not support UDP traffic and terminates connections at the Google Front End (GFE), meaning it does not preserve the client source IP address in the IP header.
• B. Incorrect. The Regional External Proxy Network Load Balancer is a proxy-based Layer 4 load balancer using the Envoy proxy. Because it terminates connections at the proxy, it does not naturally preserve the client source IP address in the packet header (requiring the PROXY protocol instead). It also introduces more latency than a passthrough solution, making it less ideal for gaming.
• C. Correct. The Regional External Passthrough Network Load Balancer is a non-proxied, Layer 4 load balancer that supports UDP traffic. Because it is passthrough, it forwards packets directly to the backend instances while preserving the original client source IP address. This architecture provides the lowest possible latency, which is a critical requirement for gaming applications.
• D. Incorrect. The Internal Application Load Balancer is a regional Layer 7 load balancer intended for internal traffic within a VPC. It is designed for HTTP(S) traffic and does not support external UDP traffic or preserve the client source IP address as it acts as a proxy.

Correct answer: A — Configure Alias IP ranges on the primary network interface.

• A. Correct. Alias IP ranges allow you to assign multiple internal IP addresses from a VPC subnet to a single network interface on a VM instance. This is specifically designed for scenarios where multiple services or containers on a single instance need unique internal IPs without the complexity and resource limits of managing multiple NICs.
• B. Incorrect. Static NAT maps internal (private) IP addresses to external (public) IP addresses for external connectivity. It does not provide additional internal IP addresses to a single interface for internal VPC communication.
• C. Incorrect. Port mapping via iptables allows multiple services to share a single IP address by using different ports. While this makes services reachable, it fails the requirement of providing each service with its own unique internal IP address.
• D. Incorrect. Shared VPC is a project-level organization tool that allows multiple projects to use a common VPC. It does not natively provide a mechanism to assign multiple internal IPs to a single instance's interface without using Alias IP ranges or additional NICs.

Correct answer: B — Create a Custom Machine Type with 6 vCPUs and 20 GB of RAM.

• A. Using a standard n1-standard-8 machine type leads to over-provisioning (paying for 2 vCPUs and 10 GB RAM that aren't needed). While Sustained Use Discounts (SUDs) provide automatic discounts for long-running workloads, they do not eliminate the base cost of the unused resources, making it less cost-effective than right-sizing.
• B. Google Cloud allows you to create Custom Machine Types where you can specify the exact number of vCPUs and amount of memory required (within valid ratios). This eliminates the cost of over-provisioned resources, ensuring you pay only for the 6 vCPUs and 20 GB RAM the application requires.
• C. An n1-standard-4 instance provides only 4 vCPUs, which is insufficient for an application requiring 6 vCPUs. Standard N1 machine types do not support burstable performance to bridge a core count gap; under-provisioning would result in performance degradation and application instability.
• D. Rewriting a proprietary legacy application to run on Cloud Run involves significant development effort, time, and cost. Furthermore, Cloud Run is optimized for stateless containers and may not be compatible with the architecture of a legacy application, making this the least efficient and most expensive migration path.

Correct answers: B, C — Configure Customer-Managed Encryption Keys (CMEK) using Cloud KMS.; Enable VPC Service Controls and define a service perimeter containing the project and GCS service.

• A. Google-managed encryption keys (GMEK) are the default encryption state for Cloud Storage. Because Google manages both the key material and the rotation schedule, this does not satisfy the requirement for the organization's security team to manage the keys or enforce a specific 90-day rotation.
• B. Customer-Managed Encryption Keys (CMEK) via Cloud KMS allow your security team to retain control over the key lifecycle, including rotation policies. Cloud KMS supports automated rotation schedules, which can be configured to meet the 90-day requirement while still being managed within the Google Cloud ecosystem.
• C. VPC Service Controls allow you to define a security perimeter around sensitive resources like Cloud Storage. This prevents data exfiltration by restricting data access and movement to only authorized services and projects within the perimeter, even if a user has valid IAM credentials.
• D. Uniform Bucket-Level Access simplifies access management by using IAM roles exclusively and disabling Access Control Lists (ACLs). While a best practice for security management, it does not provide encryption key management or data exfiltration prevention.
• E. While Customer-Supplied Encryption Keys (CSEK) offer high levels of control, they require the client to provide the key with every request and do not support automated rotation within Google Cloud. This introduces significant operational complexity compared to CMEK for meeting a strict 90-day rotation requirement.

Correct answer: C — Cloud Spanner.

• A. Cloud SQL for SQL Server is a managed relational database service suitable for typical OLTP workloads. However, it is primarily designed for vertical scaling and does not natively provide the global horizontal scaling or the massive write throughput (millions of transactions per second) required for this specific trading platform.
• B. Cloud Bigtable is a high-performance, NoSQL wide-column store designed for high throughput and low latency at massive scale. While it can handle millions of requests per second, it does not meet the requirement for a relational structure or ANSI SQL compliance.
• C. Cloud Spanner is a fully managed, horizontally scalable, relational database service that provides strong consistency, ANSI SQL compliance, and high availability across regions. It is specifically designed to handle massive transactional workloads (millions of TPS) and global replication, making it the ideal fit for a multi-region financial trading platform.
• D. BigQuery is an enterprise data warehouse designed for analytical (OLAP) processing and complex queries on large datasets. It is not intended for low-latency, high-frequency transactional (OLTP) workloads like a trading platform.

Correct answer: B — Use Vertex AI Pipelines with custom components that trigger retraining via Cloud Functions when new data is uploaded.

• A. Incorrect. Cloud Composer and Dataproc are valid for orchestration and processing, but they are not optimized for end-to-end ML lifecycle management. AI Platform is a legacy service; the modern approach uses Vertex AI. This solution lacks native automatic retriggering and integration with the ML pipeline services.
• B. Correct. Vertex AI Pipelines is the managed service for defining and orchestrating end-to-end ML workflows with reusable components for data ingestion, training, evaluation, and deployment. Cloud Functions can be used as a trigger to automatically start retraining when new data is uploaded, enabling an event-driven MLOps architecture.
• C. Incorrect. Agent Platform Pipelines is not a recognized Google Cloud service for ML workflows. While the concept is similar, the actual product is Vertex AI Pipelines. This option does not describe a valid or recommended approach for building ML pipelines on Google Cloud.
• D. Incorrect. Cloud Build is designed for CI/CD of application code and container images, not for orchestrating ML lifecycle workflows. Although it can run scheduled jobs and deploy artifacts, it lacks native support for ML pipeline components, automatic retraining triggers, and does not integrate well with the ML lifecycle management tools.

Correct answer: B — Cloud Vision API Product Search for visual similarity and matching.

• A. Incorrect. Cloud Vision API can detect objects and extract dominant colors, but it does not natively support visual similarity matching or product recommendations, as it lacks the indexed product matching capability needed for e-commerce image search.
• B. Correct. Cloud Vision API Product Search is built specifically for visual similarity search and matching against a product catalog. It enables customers to upload a reference image and find similar items using prebuilt visual search capabilities, making it the most direct solution for this use case.
• C. Incorrect. While AutoML Vision can train a custom image classification or object detection model, it is not the most direct solution for visual similarity search over a catalog. It would require additional training and setup rather than using a purpose-built product matching API, and it is not a prebuilt API but a custom training service.
• D. Incorrect. Cloud Video Intelligence API analyzes video content such as frames, labels, and shot changes, which is not relevant for static product image matching. It is not intended for product catalog similarity search.
• E. Incorrect. Vertex AI Search is designed for semantic search across structured and unstructured content using text and metadata, but it is not a specialized visual similarity API for product images. It could help query metadata but does not replace product image matching functionality.
• F. Incorrect. Cloud Storage is useful for storing and serving images, but it does not provide search, classification, or similarity matching by itself. It is only a storage layer and must be combined with another service for image search.

Correct answer: B — Configure an Organization Policy with the constraint 'constraints/storage.publicAccessPrevention' set to 'enforce'.

• A. Cloud DLP is a detective and reactive control, not a preventative organization-level policy. A weekly scan frequency allows for a significant window of exposure and does not proactively override bucket-level permissions to block public access before it happens.
• B. The 'constraints/storage.publicAccessPrevention' Organization Policy constraint is the specific preventative control designed to block public access (allUsers and allAuthenticatedUsers) at the resource hierarchy level. When enforced at the organization level, it overrides any local bucket settings, ensuring that public exposure is prevented rather than just detected after the fact.
• C. The 'constraints/storage.uniformBucketLevelAccess' constraint enforces the use of uniform bucket-level IAM policies by disabling per-object ACLs. While this simplifies security management, it does not itself prevent a user from granting public IAM bindings (like allUsers) to the bucket.
• D. Manually removing roles is a reactive, manual, and error-prone process. It is not scalable for millions of documents and does not prevent future buckets or permissions from being created with public access. It fails to meet the requirement for an organization-level override.

Correct answer: E — Configure sensitive data redaction directly in Cloud Logging using a log bucket with a custom de-identification template applied at the ingestion level.

• A. Incorrect. Routing logs to Pub/Sub and using a Dataflow job plus Cloud DLP introduces significant operational complexity, requiring the management of pipelines, topics, and worker scaling. It is not the path of least operational effort.
• B. Incorrect. While BigQuery authorized views can mask data for developers, the raw PHI is still stored in the underlying BigQuery table. The requirement explicitly states that de-identification must happen before logs are stored permanently.
• C. Incorrect. Redacting data within the application before writing to stdout/stderr requires modifying application code and manually managing DLP API calls. This increases operational effort and development overhead compared to platform-native solutions.
• D. Incorrect. This architecture involves multiple steps (Storage, Cloud Functions, and DLP), creating high operational overhead. Furthermore, the raw logs are temporarily stored in the initial bucket before the function processes them, potentially violating the 'before storage' requirement.
• E. Correct. Google Cloud Logging now supports native integration with Sensitive Data Protection (Cloud DLP) at the Log Bucket level. By configuring a de-identification template and applying it at ingestion, PHI is redacted before it is permanently stored in the log bucket, meeting the compliance requirement with the lowest possible operational effort.

Correct answer: B — Refuse to provide the name and explain the concept of a 'Blameless Post-Mortem,' focusing on fixing the process gap that allowed the error to happen rather than punishing the individual.

• A. Incorrect. Identifying individuals for reprimand or specific remedial training creates a culture of fear and blame. This discourages engineers from reporting mistakes and shifts the focus away from systemic vulnerabilities, directly contradicting SRE principles.
• B. Correct. A cornerstone of SRE is the 'Blameless Post-Mortem.' This approach assumes that human error is inevitable and focuses on identifies the systemic causes (e.g., lack of automated validation, inadequate testing environments, or poor deployment strategies) that allowed the error to impact production. This fosters psychological safety, transparency, and long-term organizational learning.
• C. Incorrect. Deflecting blame to an external vendor avoids internal accountability and misses the opportunity to improve internal processes. It does not address why the configuration drift occurred or why internal monitoring failed to catch it.
• D. Incorrect. While peer reviews are often beneficial, implementing a 'two-person rule' specifically to appease a stakeholder is a reactive measure that introduces operational friction without addressing the underlying cultural issues or technical root causes.

Correct answer: B — It shifts the model from CapEx to OpEx, eliminates the need for a dedicated Hadoop ops team, and handles scaling automatically.

• A. BigQuery primarily uses GoogleSQL (Standard SQL). While it historically offered a legacy SQL mode, it is not 100% compatible with arbitrary legacy on-premise dialects, and migrating complex queries usually requires refactoring.
• B. BigQuery is a serverless, fully managed analytics service. By choosing it over a self-managed Hadoop cluster on Compute Engine, Mountkirk shifts from a CapEx (investing in hardware/infrastructure) to an OpEx (pay-as-you-go) model. It eliminates the operational burden of managing Hadoop nodes and automatically scales to handle large workloads without manual intervention.
• C. BigQuery cannot directly read on-premise HDFS files in-place. Data must be migrated to Google Cloud (such as Cloud Storage or ingested into BigQuery storage) before analysis, meaning existing HDFS files cannot be reused directly without a transfer process.
• D. BigQuery is optimized for Online Analytical Processing (OLAP) and large-scale data analysis. It is not a low-latency transactional (OLTP) database, making it inappropriate for the core transactional processing tasks usually handled by an OLTP system.

Correct answers: A, B — Enable the Container Analysis API and configure a vulnerability scanning policy in Artifact Registry. Block deployment if vulnerabilities are found.; Use Binary Authorization with an Attestor. Configure the policy to require an attestation from the QA team before the image can be deployed.

• A. Correct. Enabling the Container Analysis API and using Artifact Registry vulnerability scanning allows you to automatically identify CVEs and severity metadata for container images. This metadata can be used to prevent deployment—often in combination with Binary Authorization—to ensure that only images meeting your security standards enter production.
• B. Correct. Binary Authorization provides a deployment-time security control that ensures only trusted container images are deployed on GKE. By creating an Attestor for the QA team, you can enforce a policy that requires a cryptographic signature (attestation) verifying that the image has passed QA inspection before it is permitted to run.
• C. Incorrect. While custom automation can detect vulnerabilities, this approach is brittle and lacks the native integration and scalability of the Container Analysis API. Deleting images is a destructive action that does not provide a formal admission control mechanism to prevent unauthorized deployments.
• D. Incorrect. Kubernetes Network Policies are used to control L3/L4 traffic between pods and services. They do not have the capability to inspect container image metadata, vulnerability status, or cryptographic signatures, which are handled at the admission control level.
• E. Incorrect. Manually checking image hashes against a custom database in Cloud Storage is not a scalable or industry-standard solution. It lacks the automated vulnerability intelligence provided by Container Analysis and the secure, policy-based enforcement provided by Binary Authorization.

Correct answer: A — Configure a Cloud Build trigger that runs a build config file (cloudbuild.yaml) to execute pytest and then gcloud app deploy upon push to the repository.

• A. Correct. Cloud Build is Google Cloud's native, serverless CI/CD platform. Configuring a trigger to run a build config file upon a repository push provides an event-driven, fully managed workflow. Using pytest for testing and gcloud app deploy for App Engine integration ensures a seamless and scalable pipeline without the need to manage infrastructure.
• B. Incorrect. While Jenkins is a robust CI/CD tool, running it on a Compute Engine instance makes it a self-managed solution. This requires the team to handle provisioning, maintenance, OS patching, and scaling, which is less integrated and managed compared to Cloud Build.
• C. Incorrect. Cloud Scheduler is designed for time-based jobs, not event-driven CI/CD. Polling a repository introduces latency, inefficiency, and unnecessary operational complexity compared to a trigger-based system like Cloud Build.
• D. Incorrect. Manual testing and manual deployment are error-prone, lack auditability, and do not satisfy the requirement for an automated process. This approach is not scalable for a modern development lifecycle.

Correct answer: C — Configure a Google Cloud Storage (GCS) backend.

• A. Storing the state file in a Git repository is discouraged because it exposes sensitive state data (like passwords and keys) to VCS history and does not provide state locking. Concurrent Terraform runs will lead to race conditions and merge conflicts, making it unsuitable for team-based or CI/CD workflows.
• B. A local backend on a shared network drive lacks the reliability, security, and fine-grained IAM controls required for enterprise environments. It does not provide the native locking mechanisms Terraform requires to prevent concurrent modifications and is error-prone when accessed by multiple team members or automated pipelines.
• C. The Google Cloud Storage (GCS) backend is the native and recommended remote backend for Terraform on GCP. It provides secure storage with encryption at rest, integrates with IAM for access control, supports object versioning for state recovery, and natively supports state locking to prevent concurrent modifications, ensuring consistency across team and CI/CD operations.
• D. Cloud SQL is not a supported Terraform remote backend. While relational databases are used by some other backends (like the 'pg' backend for Postgres), Cloud SQL is not natively integrated for state locking and management in the same way GCS is, and using it would introduce unnecessary complexity and overhead.

Correct answer: B — Visual Studio Code and IntelliJ IDEA

• A. Incorrect. Eclipse and NetBeans do not have official Cloud Code extensions. While there were legacy plugins like 'Google Cloud Tools for Eclipse', Cloud Code is a distinct, modern suite that specifically targets the VS Code and JetBrains ecosystems.
• B. Correct. Google provides official Cloud Code extensions for Visual Studio Code and IntelliJ IDEA (along with other JetBrains IDEs like PyCharm and WebStorm). These extensions integrate tools like Skaffold and minikube to streamline the development, debugging, and deployment of containerized applications to Kubernetes and Cloud Run.
• C. Incorrect. Atom and Sublime Text are popular text editors, but they do not have official first-party Cloud Code extensions provided by Google. Community-led plugins may exist for basic GCP interaction, but they lack the full Cloud Code feature set.
• D. Incorrect. Vim and Emacs do not have official Cloud Code extensions from Google. While advanced users can configure these editors to interact with Google Cloud via the gcloud CLI and API tools, they do not benefit from the specific Cloud Code graphical integrations.

Correct answer: B — Use Managed Instance Groups (MIGs) with autoscaling policies based on CPU utilization and load balancing serving capacity.

• A. Pre-provisioning a fixed number of instances is contrary to the Well-Architected Framework's principles of elasticity. It results in either underutilized resources during low traffic periods (waste) or performance degradation if traffic exceeds the fixed capacity, failing to address unpredictable spikes.
• B. Managed Instance Groups (MIGs) with autoscaling based on real-time metrics like CPU utilization or load balancing serving capacity provide the elasticity and automation required. This approach ensures the backend scales rapidly and automatically in response to actual demand, satisfying the requirement for handling unpredictable spikes without manual intervention.
• C. Scheduled scaling is only effective for predictable, cyclical traffic patterns. Since the scenario specifies that traffic spikes are unpredictable, a time-based policy would be insufficient and could lead to either resource shortages or unnecessary costs.
• D. Manual intervention based on alerts is the current slow process Mountkirk Games is trying to replace. Relying on humans to respond to alerts introduces latency and risk of error, which contradicts the Operational Excellence goal of using automation to increase speed and reliability.

Correct answers: A, C — Conduct a 'Game Day' or DiRT (Disaster Recovery Testing) exercise simulating the traffic spike and component failures.; Review and update the incident response playbooks to ensure they are current and accessible.

• A. Correct. Conducting a 'Game Day' or DiRT (Disaster Recovery Training) exercise simulates realistic traffic spikes and component failures. This validates the system's resilience and scaling behavior while providing hands-on experience for the operations team to practice incident response protocols under pressure.
• B. Incorrect. Testing in production during peak hours without warning is high-risk and can cause actual service degradation for customers. Load testing should be conducted in a controlled environment or during maintenance windows with team coordination.
• C. Correct. Reviewing and updating incident response playbooks ensures that the operations team has access to clear, documented, and tested procedures for handling failures. This reduces response time and error rates during high-pressure events like a flash sale.
• D. Incorrect. While Enterprise Support provides valuable assistance, relying solely on external support is insufficient for ensuring architectural readiness and internal operational capability. The team must still test and prepare their own systems.
• E. Incorrect. Manually scaling all resources to maximum limits a month in advance is cost-inefficient and does not address the need to test dynamic scaling mechanisms or the team's ability to handle incidents.

Correct answer: B — Initiate a manual failover from the Cloud SQL instance overview page during a maintenance window.

• A. The gcloud compute instances stop command targets Compute Engine VMs, not managed Cloud SQL instances. Attempting to stop the underlying VM of a managed service is not the correct approach to trigger a controlled failover and could lead to unpredictable results or misleading test data.
• B. Initiating a manual failover (using the Cloud SQL Console or 'gcloud sql instances failover') is the officially supported and recommended method to test High Availability configurations. This approach safely promotes the standby instance to primary, allowing you to validate that the application reconnects and handles the transition correctly without data loss.
• C. Saturating the CPU to force a health check failure is an uncontrolled and unreliable method for testing managed HA. It may cause application performance degradation or data integrity issues without cleanly triggering the failover mechanism, making it unsuitable for a formal reliability audit.
• D. Deleting the primary Cloud SQL instance is a destructive action that removes the managed resource and its metadata. This is not a failover test; it is a deletion. HA failover relies on the existence of the primary and standby within the same managed instance configuration.

Correct answer: C — Configure VirtualService fault injection in ASM to introduce aborts (HTTP 500) and delays for a percentage of traffic destined for the Inventory service.

• A. Manually deleting pods tests infrastructure availability and the platform's ability to reschedule instances, but it is a 'brute force' method. It does not allow for the granular simulation of specific application-layer error codes or high latency while the service remains reachable, making it less precise for testing service-to-service resilience.
• B. Blocking traffic at the subnet level via Traffic Director or firewall rules is a coarse-grained network action. This affects broad connectivity rather than the specific HTTP error or latency behaviors you want to simulate. Furthermore, Anthos Service Mesh (ASM) provides native, fine-grained L7 controls that are more appropriate for intra-mesh testing than network-level subnet blocks.
• C. ASM (based on Istio) natively supports VirtualService fault injection. This allows you to introduce aborts (e.g., HTTP 500 errors) and delays for a configurable percentage of traffic without modifying application code. This is the most appropriate method as it enables precise, controlled, repeatable, and reversible chaos experiments at the application layer.
• D. Modifying production code to introduce artificial failures is a high-risk anti-pattern. It mixes testing logic with business logic, is difficult to roll back quickly, and can lead to accidental production outages. Chaos engineering best practices favor using the service mesh or infrastructure layer to inject faults non-invasively.

Correct answer: B — Write a blameless post-mortem that focuses on the process failure, the timeline of events, and automated safeguards to prevent the specific misconfiguration in the future.

• A. Incorrect. This approach violates the core SRE principle of blamelessness. Identifying and punishing an individual creates a culture of fear that discourages transparency and reporting, and it ignores the systemic or process failures that allowed the human error to reach production.
• B. Correct. A blameless post-mortem focuses on systemic issues and process improvements rather than individual mistakes. By documenting the timeline, root causes, and implementing automated safeguards (such as policy-as-code or automated validation tests), the team improves system reliability and fosters a culture of continuous learning.
• C. Incorrect. Extreme centralization and manual gatekeeping by senior leadership create operational bottlenecks and do not address the technical root cause. SRE emphasizes least privilege, peer review, and automated guardrails over hierarchical restrictions.
• D. Incorrect. Suppressing documentation prevents the organization from learning from the event. SRE practices require detailed incident analysis to identify the 'what' and 'how' of a failure to ensure it does not happen again.

Correct answer: A — Network Intelligence Center Connectivity Tests

• A. Network Intelligence Center Connectivity Tests is the Google Cloud native diagnostic tool designed specifically to perform reachability analysis. It simulates the network path between two endpoints (including hybrid scenarios via Cloud VPN) and can identify whether a failure is due to routing, firewall rules, or VPN configuration errors by providing hop-by-hop diagnostics.
• B. While traceroute can show hops and help infer where traffic stops, it is a standard networking utility and not a Google Cloud native configuration analysis tool. It may be blocked by ICMP/TCP filtering and cannot inspect or explain internal GCP configuration logic like VPC firewall rules or dynamic route tables.
• C. VPC Flow Logs record accepted and rejected flows for network interfaces. While they are useful for post-hoc analysis and monitoring traffic patterns that have already reached the VPC, they do not actively simulate or trace the connectivity path from on-premises to a VM to provide immediate diagnostic insights into configuration errors.
• D. The Cloud Logging agent is used for collecting system and application logs from within VM instances. It is not a network diagnostic tool and has no visibility into the network path, routing, firewall evaluation, or VPN connectivity.

Correct answers: A, D — Deploy the new version using gcloud app deploy --version=v2 --no-promote.; Configure Traffic Splitting in the App Engine console to route 20% to v2 and 80% to v1 using 'IP address' as the shard by method.

• A. Correct. Deploying the new version using the `--no-promote` flag ensures that the version is uploaded and ready but does not automatically receive all traffic. This allows you to manually configure traffic splitting between v1 and v2 afterward.
• B. Incorrect. The `--promote` flag (which is the default) immediately routes 100% of traffic to the new version. This would prevent the marketing team's requirement of an 80/20 split.
• C. Incorrect. Splitting traffic by 'Cookie' uses a unique cookie value to assign users to versions. While this provides session persistence, it does not satisfy the specific requirement to maintain IP affinity.
• D. Correct. App Engine's native traffic splitting feature allows you to distribute traffic by weights. Selecting 'IP address' as the sharding method ensures that requests coming from the same IP address are consistently routed to the same version, meeting the IP affinity requirement.
• E. Incorrect. While Cloud Load Balancing can manage traffic, App Engine has built-in traffic splitting capabilities specifically designed for this purpose. Using an external load balancer to split between internal App Engine versions adds unnecessary complexity and is not the standard practice for this scenario.

Correct answer: B — Create a new MIG with the new version. Configure the Global External Load Balancer to split traffic 95% to the old MIG and 5% to the new MIG. Monitor custom metrics in Cloud Monitoring. If stable, gradually shift traffic.

• A. Rolling updates are typically used for gradual deployment across a single group. While setting maxUnavailable to 0 preserves capacity, it does not easily allow for routing a specific, small percentage of live traffic for side-by-side metric comparison over a period. Furthermore, rolling back requires initiating another update process, which is not as fast as shifting traffic at the load balancer level.
• B. This approach implements a Canary Deployment using weighted load balancing. By creating a separate MIG for the new version and using the Global External Load Balancer to split traffic (e.g., 95/5), you can safely test the new logic on a small subset of live users. This allows for real-time comparison of custom metrics in Cloud Monitoring. If issues are detected, you can instantly roll back by shifting 100% of the traffic back to the original MIG.
• C. Load testing in a staging environment is a best practice, but it cannot fully replicate the unpredictable nature of live production traffic and real user behavior. This strategy also lacks the live canary testing component and the instant rollback capability required by Mountkirk Games.
• D. Updating all instances simultaneously (a 'Big Bang' deployment) is high risk and does not allow for testing with a small percentage of users. Additionally, restoring from snapshots is a slow, disruptive process that involves downtime, failing the requirement for an 'instant' rollback and high reliability.

Correct answer: B — Latency (golden signal), Error rates (HTTP 5xx), and Saturation of the canary instances compared to the baseline instances.

• A. Incorrect. CPU utilization and Disk I/O of the build server relate to the CI/CD build infrastructure and pipeline efficiency, not the runtime behavior or health of the API version being deployed to production. They cannot indicate user-facing errors or performance regressions.
• B. Correct. Latency, error rates, and saturation are part of the 'Four Golden Signals' of SRE. Comparing these metrics between canary (test) and baseline (control) instances allows for a statistically significant determination of whether the new version is behaving correctly and performing optimally under real-world traffic.
• C. Incorrect. The number of lines of code changed is a static metric that might inform pre-deployment risk assessment, but it provides no information about the runtime behavior, stability, or user experience of the API once it is deployed.
• D. Incorrect. While load balancer ingress cost is a useful financial metric, it is not an indicator of service health or correctness. It does not provide the immediate feedback needed to decide whether a technical rollout should proceed or be rolled back due to failure.

Correct answer: C — Create a Log Sink to export logs to a Pub/Sub topic. Create a Pub/Sub subscription and use Dataflow or a Splunk add-on to pull messages to the on-premise system.

• A. Exporting logs to Cloud Storage is a batch-oriented process. Using a scheduled script to download files introduces significant latency and operational overhead, failing the requirement for near real-time ingestion.
• B. BigQuery is optimized for analytics rather than continuous real-time streaming to external systems. Connecting via ODBC involves polling, which adds latency, complexity, and potential egress costs, making it unsuitable for a real-time SIEM pipeline.
• C. This is the Google Cloud recommended architecture for log streaming to SIEMs. A Log Sink exports logs to a Pub/Sub topic with minimal latency. Dataflow or the Splunk Add-on for Google Cloud then retrieves these messages, providing a highly scalable, reliable, and near real-time ingestion path.
• D. Polling the Cloud Logging API is inefficient and brittle. This approach is subject to API rate limits and higher latency. It lacks the delivery guarantees and scalability of a Pub/Sub-based streaming architecture.

Correct answer: B — Google Kubernetes Engine (GKE)

• A. Compute Engine is an Infrastructure-as-a-Service (IaaS) offering for running virtual machines. While you can manually install container runtimes on VMs, it is not a dedicated management and orchestration service for containerized applications.
• B. Google Kubernetes Engine (GKE) is the standard managed service for deploying, managing, and scaling containerized applications using Kubernetes. It automates cluster management, scaling, and health checking.
• C. Cloud Functions is a serverless Function-as-a-Service (FaaS) environment for executing small, event-driven code snippets. It is not designed for managing full-scale containerized application architectures.
• D. App Engine is a Platform-as-a-Service (PaaS) offering. While the App Engine Flexible environment does use Docker containers, GKE is the primary service designed specifically for broad container management and orchestration.