Prepare and test your skills
Prepare and test your skills
Assess Regulatory Compliance and Risk Mitigation is the practice of designing cloud architectures that meet global regulations and industry standards while reducing legal, security, and operational risk. A governed architecture maps business goals and market rules to technical controls for scanning, detection, secure configuration, and safe data use. Those controls work together so risk is prevented before deployment, detected during operation, and fixed in a timely manner.
A governed architecture for global and industry requirements translates regulatory obligations into enforced technical choices across different markets. The architecture defines who owns each control, where trust boundaries sit between users, applications, and data, and how data flows from its source to its destination. For regulated workloads, the architecture combines preventive controls that harden infrastructure, detective controls that log and monitor activity, and corrective controls that remediate findings after scans or significant changes.
Continuous vulnerability scanning and remediation is the SOC 2 CC7.1 control that requires periodic infrastructure and software scans plus extra scans after significant environment changes. Artifact Analysis (opens in a new tab) vulnerability scanning identifies vulnerabilities or misconfigurations in containers and artifacts so deficiencies can be remediated in a timely manner to support business objectives. Scan results flow from the scanner to the remediation process, and remediation is complete only after the identified deficiency is fixed and verified.
Detection controls for intrusions and anomalous activity are the SOC 2 CC7.2 policies, procedures, and tools that identify intrusions, inappropriate access, and unusual system activity. The detection system combines governance for security event management, intelligence sources for new threats, and logging of unusual activities to keep visibility across infrastructure and software. These controls lower operational risk by making misconfigurations and suspicious behavior visible before they cause legal or security failures.
The supporting Google Cloud hardening controls protect compute, network, and database layers:
Exam tip: Use Artifact Analysis vulnerability scanning for periodic scans and scans after major changes, while Shielded VM, TLS enforcement, and database logging support continuous detection of intrusions and anomalies.
Secure retrieval for retrieval-augmented generation (RAG) applications on SAP enterprise data is a serving pattern that grounds large language model responses in enterprise facts. The serving subsystem in SAP manages the full request flow, while the ABAP SDK for Google Cloud (opens in a new tab) handles communication between the subsystem and Google Cloud services. Data flows from the SAP data source through the serving subsystem to Agent Platform Gemini models, which generate responses grounded in enterprise context for improved accuracy and user experience.
The serving subsystem processes each user request in a fixed order:
A process flow showing the SAP RAG serving workflow: a user request flows through the frontend to the serving subsystem, which retrieves SAP data via a function module or Vector Search, augments the prompt through the ABAP SDK, sends it to Agent Platform Gemini models, and delivers the grounded response back to the user.
The retrieval method depends on data size and latency needs:
| Retrieval approach | When to choose it | How data flows |
|---|---|---|
| SAP function module | Dataset is small and resides within SAP systems | Data flows from SAP tables through SELECT queries, BAPI calls, or Gemini function calling to the serving subsystem |
| Vector Search products | Enterprise data is large and low latency is required | Embeddings flow from sources to a vector database such as Cloud Storage, Agent Platform Feature Store, or BigQuery, then semantic search results flow to the serving subsystem |
Risk mitigation for AI workloads and service communication layers application, network, and mesh controls around prompts, responses, and service-to-service traffic. Model Armor inspects and sanitizes prompts, tool interactions, and responses to protect against prompt injection, sensitive data leaks, and harmful content. Google Cloud Armor enforces web application firewall (WAF) rules on incoming requests, while Identity-Aware Proxy (IAP) enforces a zero-trust boundary by verifying user identity before traffic reaches the application.
Cloud Service Mesh (opens in a new tab) extends governance to microservice communication without changing application code. The control plane manages traffic policy and telemetry, while the data plane carries traffic between services through sidecar proxies on Kubernetes or proxies and proxyless gRPC on Compute Engine. Traffic flows from ingress into the mesh, between named services and their endpoints such as Pod IPs or managed instance group VM IPs, and out to egress destinations under centrally managed routing, load balancing, and security policy.
Align Technical Roadmaps with Product Strategy is the discipline of choosing cloud building blocks that advance business goals such as better accuracy, better user experience, faster time-to-market, and global scalability. The serving subsystem in SAP for retrieval-augmented generation shows this alignment by turning enterprise SAP data into useful context for generative answers. Requests flow from users to data sources and large language models and back again, so product goals and technical design stay linked.
The serving subsystem as a product-aligned foundation is the component that retrieves relevant information, combines it with a prompt, interacts with generative models, and delivers the final response to the user. Requests flow from the application frontend to the serving subsystem, then from the serving subsystem through the ABAP SDK for Google Cloud (opens in a new tab) to Google Cloud services, and finally back from the models to the user as a grounded response. Combining Gemini large language models with SAP enterprise data produces more accurate, enterprise-grounded decisions and more personalized, contextually relevant experiences that improve satisfaction.
The architecture divides responsibilities so each part can evolve with the product roadmap without rebuilding the whole system:
Retrieval choices that support scale and speed are the two information retrieval paths offered at the first stage of the serving subsystem. Information retrieval fetches relevant enterprise context after a user submits a request through the frontend. Choosing the right path lets a team launch quickly on existing SAP data and then scale globally as data volume and performance demands grow.
The retrieval decision depends on data size, location, and latency needs:
| Retrieval path | Best when | How it works |
|---|---|---|
| SAP function module | Dataset is small and resides within SAP systems | Data flows from SAP tables and logic to the serving subsystem through SELECT queries, BAPI calls, or SAP function calling with Gemini |
| Vector Search (opens in a new tab) | Enterprise data is large and the RAG application needs minimal latency | Data flows from embeddings in a vector database to the serving subsystem through semantic search, including text and multimodal search over billions of records within milliseconds |
Once the retrieval path is selected, the roadmap can expand without changing the overall flow. A team can start with function-module retrieval for fast time-to-market, then add Vector Search storage in Cloud Storage, Agent Platform Feature Store, or BigQuery as volume grows. Because the ABAP SDK for Google Cloud depends on Google Cloud services for search and generation, the serving subsystem depends on the SDK to keep SAP logic and cloud services connected during each phase.
Exam tip: Choose SAP function modules when data is small and inside SAP; choose Vector Search when data is large and needs low-latency semantic search.
The augmentation and generation workflow that preserves context covers the second and third stages of the serving subsystem. Information augmentation combines the retrieved enterprise facts with the user's prompt to create model-ready context. The augmented prompt flows from the serving subsystem to the Agent Platform Gemini models (opens in a new tab), so the model receives both the original question and the supporting SAP data.
Response generation is the final stage, where the Gemini models generate output grounded in the augmented context and the serving subsystem delivers it back to the user. The ordered workflow preserves business context end to end:
This sequence keeps product goals and technical design linked over time. Accuracy improves because a wider range of enterprise information reaches the model, and user experience improves because responses stay personalized and relevant. The same flow also supports future scale, since larger embedding stores and low-latency search can be added behind the retrieval step without changing how augmentation and generation work.
Evaluate Business Value and Financial Impact is the analysis of what an architecture costs over time compared with the business outcomes it protects. The cloud bursting pattern (opens in a new tab) shows this tradeoff by keeping steady baseline load in a private on-premises environment and temporarily bursting to Google Cloud only when extra capacity is needed. This approach lowers Total Cost of Ownership (TCO) by avoiding permanent overprovisioning while improving Return on Investment (ROI) by preventing service interruptions during peak demand.
The cloud bursting pattern for baseline and peak demand is a redundant deployment where the same application runs across multiple computing environments to increase capacity, resiliency, or both. Baseline traffic and batch or CI/CD jobs run in the private on-premises environment, and data and requests flow from that on-premises environment to Google Cloud in burst mode when local capacity reaches its limit. The pattern applies to both interactive workloads and batch workloads, including enterprise batch jobs that do not face internet-scale swings but still spike sharply.
The financial case rests on saving money and reducing the time and effort needed to respond to scale changes. The organization pays only for the cloud resources used while handling extra load, so it does not carry idle servers year-round for short peaks. On-demand cloud resources scale up and down to fit demand and predefined metrics, which avoids both wasted capital and lost revenue from interruptions during peak demand. Long-term financial sustainability comes from paying for on-demand scale that follows actual demand, since overprovisioning a data center for spikes is often not cost effective and stretching time-sensitive batch jobs over longer periods is not practical.
Distributing interactive workloads across environments means deciding how incoming requests flow from clients to backends located both on-premises and in Google Cloud. The load balancer sits at the trust and network boundary between the private environment and Google Cloud, and the choice determines who tracks cloud capacity, who triggers scaling, and how much hybrid connectivity must carry. The supported options differ in control, complexity, and added security benefits.
| Option | How request flow works | Benefit | Tradeoff |
|---|---|---|---|
| On-premises load balancer | Client requests flow to a load balancer in the data center, then the load balancer distributes requests across local and cloud resources | All cloud resources can be decommissioned during low activity | The on-premises system must track cloud resources and trigger upscaling and downscaling, which can exceed load balancer capabilities and increase complexity |
| Cloud Load Balancing (opens in a new tab) with hybrid connectivity network endpoint group (NEG) | Client requests flow to Cloud Load Balancing, then the load balancer routes to on-premises and Google Cloud backends using metrics such as weight-based traffic splitting | Backends in Google Cloud can scale on load balancing serving capacity, plus traffic can use Google Cloud Armor DDoS protection, WAF, and edge caching with Cloud CDN | Hybrid network connectivity must be sized to handle the additional burst traffic |
Choose the on-premises load balancer when the priority is fully shutting down cloud spend between bursts and the team can operate the tracking and autoscaling logic. Choose Cloud Load Balancing with hybrid connectivity NEG when the priority is simpler scaling and stronger edge protection, and the business accepts the cost of larger hybrid connectivity to carry burst traffic.
Exam tip: An on-premises load balancer can decommission all cloud resources during low activity but increases complexity, while Cloud Load Balancing with a hybrid NEG simplifies scaling and adds DDoS, WAF, and caching benefits but requires sized hybrid connectivity.
Workload portability and performance expectations define whether a workload can deploy to multiple environments without being rewritten for each infrastructure. Portability depends on abstracting away differences between the private environment and Google Cloud, and Kubernetes (opens in a new tab) provides that consistency at the workload level across diverse infrastructures. Google Kubernetes Engine (GKE) hybrid environments are the reference approach for achieving that consistent deployment model. A portable application does not perform equally in both environments, so testing is required because underlying compute, infrastructure security, networking, and proximity to dependent services determine actual performance.
Use SAP function module retrieval for small datasets that reside within SAP systems and use Vector Search retrieval for large enterprise data that requires low latency. Data flows from SAP tables through SELECT queries, BAPI calls, or Gemini function calling to the serving subsystem for function module retrieval, while embeddings flow from sources to a vector database such as Cloud Storage, Agent Platform Feature Store, or BigQuery and semantic search results flow to the serving subsystem for Vector Search retrieval.
An on-premises load balancer receives client requests in the data center and distributes them across local and cloud resources, allowing all cloud resources to be decommissioned during low activity at the cost of requiring the on-premises system to track cloud resources and trigger upscaling and downscaling with greater complexity. Cloud Load Balancing with hybrid connectivity network endpoint group (NEG) receives client requests in Google Cloud and routes them to on-premises and Google Cloud backends using metrics such as weight-based traffic splitting, allowing backends to scale on serving capacity with Google Cloud Armor DDoS protection, web application firewall (WAF), and Cloud CDN edge caching while requiring hybrid network connectivity to be sized to handle burst traffic.
Artifact Analysis vulnerability scanning identifies vulnerabilities or misconfigurations in containers and artifacts to support periodic infrastructure and software scans plus extra scans after significant environment changes. Scan results flow from the scanner to the remediation process, and remediation is complete only after the identified deficiency is fixed and verified.
An enterprise is planning to migrate its on-premises analytics workloads and applications to Google Cloud. Executive leadership requires an architectural and operational framework that ensures long-term financial sustainability and clearly demonstrates business value.
The enterprise has specified several core requirements:
Which strategy should the Cloud Architect recommend to satisfy these business and financial requirements?
Last updated