Prepare and test your skills
Prepare and test your skills
Business use cases and product strategy is the practice of linking business goals, market rules, and financial limits to specific cloud architecture choices. A governed design maps regulations and product goals to preventive, detective, and corrective controls, while product-aligned patterns such as retrieval-augmented generation (RAG) and cloud bursting turn enterprise data and elastic capacity into business value.
Governed controls for compliance and risk translate global and industry obligations into enforced technical choices for scanning, detection, and secure configuration. Preventive controls harden compute, network, and database layers, detective controls log and monitor activity, and corrective controls remediate findings after scans or major changes. Artifact Analysis supports SOC 2 CC7.1 with periodic vulnerability scans of containers and artifacts plus extra scans after significant changes, with remediation complete only after deficiencies are fixed and verified. Detection for SOC 2 CC7.2 combines security event governance, threat intelligence, and logging of unusual activity, supported by Shielded VM and Shielded GKE Nodes with Secure Boot, minimum TLS 1.2, restricted default networks, and hardened AlloyDB and Cloud SQL logging flags. Data flows from scanners and logs to remediation and review processes so risk is prevented before deployment and detected during operation.
Product-aligned retrieval for SAP enterprise data is a serving pattern that grounds generative answers in enterprise facts to improve accuracy and user experience. The serving subsystem in SAP manages the full request flow, while the ABAP SDK for Google Cloud connects SAP logic to Google Cloud retrieval and generation services. Data flows from the SAP source through the serving subsystem to Agent Platform Gemini models, which generate a response grounded in enterprise context. The subsystem runs in fixed order: first information retrieval after the frontend forwards the user request, then information augmentation that merges retrieved facts with the prompt, then response generation that sends the augmented prompt to the model and delivers the answer back to the user.
The retrieval choice 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 | 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 |
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.
Cloud bursting for baseline and peak demand keeps steady load in a private on-premises environment and temporarily bursts to Google Cloud when extra capacity is needed. Baseline traffic and batch or CI/CD jobs run on-premises, and data and requests flow from on-premises to Google Cloud in burst mode when local capacity reaches its limit. The organization pays only for cloud resources used during the burst, which lowers Total Cost of Ownership by avoiding idle overprovisioning while protecting availability during peaks. Kubernetes and Google Kubernetes Engine hybrid environments provide workload portability across infrastructures, but performance still differs by compute, security, networking, and proximity to dependencies, so testing is required.
Continue with this study guide: Business use cases and product strategy
Identifying functional and non-functional requirements turns business objectives into what the system must do and how well it must behave. Functional requirements define organization setup, product readiness, hosting location, and data movement, while non-functional requirements define availability, scalability, performance, security, residency, and cost limits.
Quality attributes, indicators, and objectives translate resilience and speed goals into measurable targets. Availability requires continued operation during hardware, software, or network failure, scalability requires automatic growth during spikes, and performance requires low latency and sustained throughput. A service level indicator measures behavior from the user perspective, such as percentile latency or successful-response ratio, while a Service Level Objective sets the target with a bound and time window, such as 95th percentile latency below 300 ms over a rolling 30-day period. Google Cloud supports these attributes with redundancy of everything, geographically distributed data centers, automatic shifting of control planes from a failed facility to a healthy facility, and automatic mirroring of data across locations. Recovery is judged against Recovery Time Objective for allowed downtime and Recovery Point Objective for allowed data loss.
Marketplace listing and hosting patterns define the functional entry conditions for selling a product on Google Cloud Marketplace (opens in a new tab). The partner organization must maintain good standing in Google Cloud Partner Network, be incorporated in a supported payment region, and maintain a vendor account and payment profile in good standing. The product must be production-ready and not in alpha or beta, must be enterprise-ready with support and strong security and no known malicious code, must host software primarily on Google Cloud, and must drive meaningful Google Cloud consumption. Hosting patterns specify where compute, data, control, and AI agents run and which side grows fastest with usage, including fully hosted on Google Cloud, compute or data plane on Google Cloud, storage and backup on Google Cloud, migration to Google Cloud, agent data analysis on Google Cloud, datasets hosted on Google Cloud, deployment on Google Distributed Cloud, and AI agents hosted on or hybrid with Google Cloud.
Exam tip: Pattern 2 and Pattern 9 share the same decision rule — the Google Cloud-hosted compute, data plane, or agent must grow faster than any off-cloud component when usage grows.
Security, residency, and budget constraints turn laws, policies, and funding limits into non-functional requirements. The shared responsibility model (opens in a new tab) assigns network and physical infrastructure to Google Cloud and access policies and data to the customer, with default encryption inherited as audit evidence. Industry frameworks include PCI DSS for payments, HIPAA for electronic personal health information, and FedRAMP (opens in a new tab) for U.S. federal workloads with a Customer Responsibility Matrix, System Security Plan, and service-specific guides. Location rules map each dataset to an allowed region, such as keeping EU customer data in EU regions for GDPR.
| Authorization level | Control baseline | Data boundary requirement |
|---|---|---|
| FedRAMP High | Full High baseline including data residency control SA-9(5) | Must use Assured Workloads Data Boundary for FedRAMP High |
| FedRAMP Moderate | Subset of High baseline with fewer controls | Can use any FedRAMP High authorized service inside the Moderate boundary |
Customer-owned protections include strong authentication, context-aware access, Binary Authorization for Borg for code approval, Cloud Audit Logs, Access Transparency, and infrastructure design (opens in a new tab) with ALTS for internal RPC traffic. Budget work connects each initiative to objectives and key results and optimizes at the system level rather than only within silos.
Continue with this study guide: Identifying functional and non-functional requirements
Business continuity planning translates tolerance for downtime and data loss into disaster recovery architectures that preserve operations and data integrity. Recovery Time Objective measures allowed downtime before service is restored, while Recovery Point Objective measures allowed data loss, and smaller values require more ready infrastructure and more frequent replication.
Recovery objectives and standby readiness define how much recovery capacity runs before a disaster. RTO is controlled by readiness of infrastructure, software, and data at the recovery site, while RPO is controlled by backup frequency and retention. Tier 1 systems such as payment storefronts may require zero RTO and zero RPO, Tier 2 systems such as CRM may allow minutes to an hour, and Tier 3 back-office systems may allow hours. Business-critical operations must depend only on data plane components that continuously process operations, not on management plane operations such as VM-creation APIs that may be unavailable during the outage. High availability typically uses synchronous replication across zones in one region for near-zero RTO and RPO, while disaster recovery typically uses asynchronous replication across regions with defined failover and replication lag.
| Pattern | What runs before failover | Recovery and cost tradeoff |
|---|---|---|
| Cold standby | Backups in Cloud Storage, images, and templates; no running servers | Largest RTO and RPO, lowest cost |
| Warm standby | Minimal database runs and receives replication; app servers restored from snapshots | Small RTO and RPO, medium cost |
| Hot standby | Full secondary environment runs and serves traffic with continuous replication | Near-zero RTO and RPO, highest cost |
Traffic steering and ordered failover keep service continuity by directing clients to healthy capacity during zonal or regional outages. Traffic flows from clients into Cloud Load Balancing (opens in a new tab), which steers to healthy backends based on health checks, while Cloud DNS redirects requests between environments because DNS zones are global resources resilient to zone and region outages. A global load balancer survives zone and region outages but can become a single point of failure across all geographies if misconfigured, while a regional load balancer survives zone outages but fails with its region. Defense in depth keeps regional load balancers as backups behind Cloud DNS so traffic flows from clients to the nearest regional balancer if the global front end is unavailable. A typical cold failover runs in order: first create the recovery deployment from the template, then restore the latest backup and transaction logs, then validate with simulated user scenarios, then switch Cloud DNS to the recovery web server. Failback reverses the flow only after the primary is restored, with data copied back first, recovery traffic stopped, DNS switched, and the recovery deployment deleted.
Exam tip: High availability means synchronous replication across zones in one region, while disaster recovery means asynchronous replication across regions where failover steps and replication lag affect RTO and RPO.
Data resiliency and deletion protection keep data persistent during zone outages, regional disasters, and corruption events. Cloud Storage dual-region buckets store objects in a chosen region pair and multi-region buckets store objects in at least two areas in the multi-region, with 99.99% availability and asynchronous synchronization between areas. Turbo replication replicates 100% of newly written objects between both regions of a dual-region pair within 15 minutes for a 15-minute RPO. Firestore in Datastore mode regional databases synchronously replicate across at least three zones and multi-region databases synchronously replicate across five zones in three regions with two serving regions and one witness region. BigQuery spreads storage and compute across zones but needs cross-region replication or exports for regional loss, Bigtable production instances use at least two clusters in unique zones with automatic replication, and Filestore Regional instances replicate synchronously across three zones. Operational recovery uses object holds, soft delete with a seven-day default, and Object Versioning, while compliance immutability uses Bucket Lock and Object Retention Lock. Backup and DR Service backs up Compute Engine Persistent Disk snapshots to Cloud Storage vaults, and Storage Transfer Service and Transfer Appliance move data for backup and migration.
Continue with this study guide: Business continuity plan
Cost optimization matches cloud capacity and pricing to real demand so spending follows business value. Recommender suggests rightsizing from utilization history, Cloud Monitoring supplies the metrics behind those suggestions, and Cloud Billing reports show the financial effect, with hierarchy and labels tying every dollar to its owner.
Storage classes and automated lifecycle keep each object in the cheapest storage that still meets access and retention needs. New ingestion data that arrives often starts in Standard for frequent, performance-sensitive reads, older or rarely read data moves to Nearline or Coldline, and long-term retention moves to Archive. Object Lifecycle Management pairs a condition such as object age with an action, so data flows from Standard to a cheaper class and finally to deletion only after the condition is met, including time-to-live deletion of logs and temporary artifacts. Autoclass instead moves each object automatically based on observed access telemetry without preset rules. Organization Policy Service enforces retention tags so lifecycle automation, FinOps reports, and carbon reports share the same owner, purpose, and retention metadata.
Exam tip: Use Object Lifecycle Management when you can define conditions for downgrade or deletion, and use Autoclass when you want Cloud Storage to transition each object automatically based on its access pattern.
Commitment discounts and flexible compute lower the billing rate through predictable commitments or use of spare capacity. Enterprise discounts give negotiated rates for minimum total spend, with Committed Use Discounts able to add further savings on top. Resource-based Committed Use Discounts commit to a minimum quantity of Compute Engine resources in a specific region for one year or three years, while spend-based discounts commit to a minimum spend on a product at the billing account level. Resource-based sharing lets combined usage from many projects consume the same regional commitment. Spot VMs use unused capacity at more than 80% discount but can be preempted, so they suit fault-tolerant batch jobs and not tiers requiring high availability. BigQuery edition pricing lowers cost per slot when slot use is large and steady, and region selection must balance price with latency, compliance, and sustainability.
Rightsizing, reporting, and allocation prevent over-provisioning that wastes money and under-provisioning that hurts performance. Utilization data flows from Cloud Monitoring to Recommender, and Recommender sends idle VM, unattached disk and IP, and machine-type suggestions to administrators without charge or setup. Rightsizing adjusts Compute Engine machine types and counts to observed CPU, memory, and traffic, with custom metrics and alerts triggering human-approved resizes. Billing data flows from Cloud Billing (opens in a new tab) to built-in reports and to BigQuery through billing export for label-based analysis, with budgets, anomaly detection, and FinOps hub centralizing oversight. Cost allocation attaches hierarchy from organization to folders to projects to resources plus labels such as team=banking_appdev so spend becomes accountable ownership enforced through Terraform and simple, valid conventions.
Continue with this study guide: Cost optimization
Supporting the application design uses managed integration, asynchronous patterns, and automated delivery to keep applications scalable, resilient, and safe to change. Apigee controls edge API exposure, Cloud Service Mesh controls service-to-service traffic, and Cloud Load Balancing distributes traffic to healthy capacity.
Edge exposure and mesh communication separate API policy at the boundary from routing between services. Traffic flows from clients or a cloud frontend through Apigee to backend systems, where Apigee applies access controls and traffic inspection before requests reach protected data, with Apigee hybrid supporting end-to-end mutual TLS across hybrid deployments. Requests flow directly from one service to another inside the boundary, but Cloud Service Mesh controls routing, retries, and failover so developers do not embed networking in code, with retries, failovers, and circuit breakers protecting overloaded services. Traffic flows from internet clients to a frontend load balancer, then through an internal load balancer to the backend tier, then to Cloud SQL or Spanner, so presentation, application, and data tiers scale independently. A global external Application Load Balancer spans regions, while regional deployments keep components in one region for lower latency.
Decoupled services and shared data let components continue work even when another component is slow or unavailable. Stateless services handle each request independent of previous requests so autoscalers can create and destroy instances without losing in-flight data, while stateful services are explicitly separated for careful planning. A queue sits between a producing service and a consuming service so the producer writes a message and continues while the consumer processes it later, and caching reduces database load and latency. Protection patterns include circuit breakers that stop traffic to a failing service, exponential backoff that retries with increasing delays, and graceful degradation that returns limited or cached results.
| Pattern | How data flows | Consistency and conflict handling |
|---|---|---|
| Partitioned data | Application writes twice to separate partitions with no direct database connection | Each partition stays independent with no automatic synchronization |
| Asynchronously replicated database | Updates flow from source to other database after a delay | Eventually consistent with conflict resolution for concurrent writes |
| Transactionally synchronized database | Every update flows immediately to all databases as one distributed database | Immediately consistent in all locations |
Build, delivery, and safe change move code from repository to production without rebuilding or manual handoffs. Source flows to Cloud Build, built artifacts flow to Artifact Registry for versioned storage, and Cloud Deploy promotes the same stored artifact through target environments in order to GKE, Cloud Run, App Engine flexible environment, and Cloud Run functions. Progressive infrastructure rollout deploys first to a few VMs in one zone, then monitors effects, then reverts quickly on issue, then diagnoses outside production, then expands verified changes. Global resources such as VPC networks and global load balancers are resilient to zone outages but can become single points of failure, so their changes need extra review. Automation with Terraform, Eventarc or Workflows from audit logs, autoscaling, and platform guardrails reduces human error, while managed instance groups recreate failed VMs, load balancing spreads traffic, and regional disks keep data available.
Continue with this study guide: Supporting the application design
Integration patterns with external systems connect Google Cloud to on-premises, edge, and other clouds with secure, scalable, and resilient exchange. Synchronous integration with Apigee waits for an immediate response, while asynchronous integration with Pub/Sub decouples sender and receiver so work survives spikes and outages.
Immediate APIs and decoupled messaging differ in coupling and timing. Traffic flows from the client application to the Apigee gateway and then through hybrid connectivity to the backend, with the gateway providing OAuth 2.0, TLS, rate limiting, threat protection, audit trails, and a stable facade that shields clients from backend changes. Combining the gateway with Cloud Load Balancing adds global routing, multi-region failover, Cloud CDN caching, and Cloud Armor DDoS and WAF protection. Data flows from the producing system to a Pub/Sub topic and then to each subscribing system on its own schedule, so producers and consumers do not need to be online together. Edge and IoT designs compute locally and synchronize to the cloud asynchronously to reduce transfer and processing costs. Synchronous dependencies between environments should be minimized because they couple availability and speed to both sides and the network path at the same time.
Exam tip: Minimize synchronous dependencies between Google Cloud and other environments; keep synchronous APIs for immediate responses and use Pub/Sub decoupling when traffic spikes or intermittent connectivity could slow performance or reduce availability.
Private connectivity between environments carries traffic between data centers, VPC networks, and other clouds with different bandwidth, latency, cost, and SLA tradeoffs. Internet transfer uses public IPs over the internet with no SLA and needs TLS for sensitive traffic. Cloud VPN provides encrypted IPsec tunnels over the internet with Cloud Router exchanging BGP routes dynamically. Cloud Interconnect provides private transport that does not traverse the public internet with deterministic performance and an SLA, using VLAN attachments and Cloud Router for dynamic routes.
| Dimension | Dedicated Interconnect | Partner Interconnect |
|---|---|---|
| Path | Direct between on-premises and Google network | Through a service provider to Google network |
| Setup | Requires routing equipment in a colocation facility | Uses provider connectivity, no colocation equipment |
| Bandwidth | 10-Gbps, 100-Gbps, or 400-Gbps circuits up to 8 per connection; VLANs 50 Mbps to 50 Gbps | 50 Mbps to 50 Gbps per VLAN, limited by provider |
| SLA | End-to-end SLA for Google-defined topologies | SLA between Google and provider; end-to-end depends on provider |
Cross-Cloud Interconnect provisions a managed private pair, one primary and one redundant connection at 10 Gbps or 100 Gbps, between Google Cloud and AWS, Azure, Oracle Cloud, or Alibaba Cloud for 99.9% or 99.99% high availability. Network Connectivity Center provides hub-and-spoke orchestration where VPC and hybrid spokes attach to a central hub for route exchange.
Federated access and handover exchange give external users and workloads verified identity and least-privilege access without long-lived keys. Workforce Identity Federation authenticates people from an existing identity provider, while Workload Identity Federation exchanges an external workload credential for short-lived Google Cloud credentials. Application Default Credentials let the same code use developer credentials locally and attached service account credentials on Compute Engine. VPC Service Controls defines a service perimeter around projects, networks, and services such as Cloud Storage, BigQuery, and Pub/Sub that blocks requests from unauthorized networks or identities even if IAM would allow them, with extension over authorized VPN or Interconnect for trusted hybrid callers. The handover pattern (opens in a new tab) shares data through storage rather than direct database access in ordered steps: first external workloads upload to Cloud Storage or Pub/Sub, then VPC workloads process with analytics and AI, then frontends serve results through load balancing without exposing VMs directly.
Continue with this study guide: Integration patterns with external systems
Movement of data covers large one-time migration, continuous streaming, and secure transit between on-premises, edge, and Google Cloud. The tool choice depends on data volume, bandwidth, timeline, cost, security, and whether the move can run without disrupting operations.
Bulk online and offline transfers move large datasets into Cloud Storage with different network dependence. Storage Transfer Service is a fully managed online copy over the network optimized above 1 TiB and able to scale to petabytes, depending on allocated bandwidth with parallelization, retries, TLS 1.3, and integrity validation. Transfer Appliance is offline hardware loaded locally and shipped to Google, recommended above 20 TB up to 1 petabyte to avoid bulk network use and business disruption. Planning flows in order: first identify what datasets to move together, then filter for what can move under regulatory and security rules, then choose the destination bucket, region, structure, and access policies, then assess time, cost, people, and bandwidth. Seed plus sync transfers reduce downtime by copying bulk data while the source stays open, then copying only changes after the source is locked.
| Option | Transfer method | Best-fit volume |
|---|---|---|
| Storage Transfer Service | Fully managed online copy to Cloud Storage | Optimized above 1 TiB, scales to petabytes |
| Transfer Appliance | Offline hardware loaded locally, then shipped | Recommended above 20 TB up to 1 petabyte |
Exam tip: Use gcloud below 1 TB, Storage Transfer Service above 1 TiB and for petabyte-scale online moves, and Transfer Appliance for on-premises moves above 20 TB up to 1 petabyte when bandwidth or timeline rules out network transfer.
Continuous streaming with messaging and processing ingests events as they occur without waiting for a batch window. Data flows from publishers to a Pub/Sub topic, then from the topic to Dataflow workers, then from Dataflow to sinks such as BigQuery or Cloud Storage. Pub/Sub decouples producers from processors for low-latency delivery, with default encryption, customer-managed encryption keys for key control, message storage policies for residency, publisher flow control, retry settings, and per-project throughput quotas. Dataflow runs pipeline stages in parallel across workers that depend on Compute Engine quota to start and autoscale, following ingest, then transform with filtering and enrichment, then deliver to sinks. Calls to external systems from transforms should batch elements and limit concurrency because parallel workers can overwhelm the external service. Managed sources and sinks simplify capacity planning, but Dataflow, Compute Engine, and Pub/Sub quotas must be monitored together, with separate projects isolating quota when maximum throughput is needed.
Dedicated links, VPN, and dynamic routing balance bandwidth, latency, cost, and encryption for hybrid and multicloud transit. Dedicated Interconnect provides direct physical circuits of 10-Gbps, 100-Gbps, or 400-Gbps up to 8 per connection in a colocation facility with VLANs from 50 Mbps to 50 Gbps and an end-to-end SLA. Partner Interconnect connects through a provider with VLANs from 50 Mbps to 50 Gbps and no colocation hardware. Cloud VPN creates encrypted IPsec tunnels over the internet, while HA VPN over Cloud Interconnect and MACsec for Cloud Interconnect add network-layer encryption because Interconnect alone provides private transport but does not encrypt by default. Cross-Cloud Interconnect provides managed 10-Gbps or 100-Gbps pairs to other providers without customer cross-connects. Cloud Router acts as a distributed BGP speaker that exchanges routes between on-premises, other clouds, and Google Cloud, so the VPC network depends on Cloud Router for automatic path updates when topologies change.
Continue with this study guide: Movement of data
Design decision trade-offs balance control against simplicity, cost against resilience, and consistency against latency. Fully managed serverless options minimize operations and speed delivery, while infrastructure options return low-level tuning at higher management cost.
Serverless simplicity and bare metal control differ in who manages servers and scaling. Requests flow to a Google-managed control plane that allocates compute and scales automatically, with no servers for the customer to patch. App Engine builds scalable applications without managing infrastructure, Cloud Run runs portable containerized applications, and Cloud Run functions runs event-driven code with pay-as-you-go billing. Bare Metal Solution provides sole-tenancy bare-metal servers in regional extensions connected through Partner Interconnect with a low-latency fabric, where traffic flows from the customer VPC into the extension and dedicated servers reach Google services through private API access. Google Cloud owns facilities, power, network, security, monitoring, and hardware provisioning plus initial OS or hypervisor install, while the customer owns data security, software and database lifecycle, OS and hypervisor patching, licensing, logging, and high-availability and disaster recovery design.
| Option | Who manages servers and scaling | Customer control |
|---|---|---|
| Cloud Run functions | Google Cloud manages all servers | Function code and event bindings only |
| Cloud Run and App Engine | Google Cloud manages platform and availability | Application code, container image, and configuration |
| Bare Metal Solution | Google Cloud manages facilities and hardware | Operating system, hypervisor, clusters, patching, and licensing |
Compute, storage, and availability economics trade extra replicas, geographic spread, and reserved capacity for higher cost. Compute Engine VMs offer the most configuration flexibility, GKE automates container deployment and scaling, and Cloud Run minimizes management. Stateless managed instance groups add VMs when traffic rises and remove them when demand falls, while stateful groups cannot autoscale. Spot VMs match regular VM performance at much lower cost but can be preempted, so critical tiers need regular VMs with reservations. Zonal Persistent Disk keeps data in one zone at lower cost, while regional Persistent Disk and Hyperdisk Balanced High Availability replicate synchronously between two zones, and Hyperdisk Balanced allows separate provisioning of IOPS and throughput. Regional Cloud Storage buckets replicate synchronously across zones, dual-region and multi-region buckets store redundantly in at least two locations with synchronous metadata and asynchronous data, and Filestore Regional replicates across three zones. Spanner multi-region instances and Cloud SQL high availability provide managed failover, while self-managed databases on VMs require manual updates, backups, and observer-based failover.
Replication boundaries for consistency and speed define which copies confirm a write and how far a read must travel. Synchronous replication within a region provides strong consistency with zero or near-zero RPO in an active-active topology, including regional Persistent Disk between two zones, Filestore Regional across three zones, and regional Cloud Storage buckets across zones. Single-region placement near users lowers latency and avoids cross-region charges, with Cloud DNS and global or regional load balancers routing to healthy capacity. Global scale with low-latency reads usually uses asynchronous replication where the primary acknowledges before distant replicas receive the write, so RPO is greater than zero. Spanner is the exception that combines global distribution with strong consistency through four read-write replicas across two regions plus a witness in a third region. Cloud SQL high availability provides automatic failover with RPO close to zero without changing the endpoint IP.
Exam tip: Choose Spanner for globally consistent relational data across regions, while Cloud Storage dual-region buckets with turbo replication mean metadata is synchronous but object data lags with a 15-minute RPO.
Continue with this study guide: Design decision trade-offs
Workload disposition strategies decide whether to build custom solutions, buy managed services, modify existing architectures, or deprecate legacy systems. The decision weighs competitive advantage, total cost of ownership, speed to market, and who carries reliability and maintenance work.
Modernization to reduce management effort replaces self-managed infrastructure with managed services that provide SLAs, backups, and scaling. Compute Engine managed instance groups automatically recreate failed VMs and autoscale stateless workloads, with regional groups resilient to zone outages. Google Kubernetes Engine automates container deployment and scaling, with regional clusters also resilient to zone outages. Cloud Run runs code on a preconfigured serverless platform with minimal operation. Modern data management replaces databases on VMs with Cloud SQL high availability and read replicas, Spanner multi-regional synchronization for active-active topologies, and Cloud Storage regional, dual-region, and multi-region buckets. Regional persistent disks replicate synchronously across two zones, Filestore Regional replicates across three zones for shared files, and turbo replication gives dual-region buckets a 15-minute RPO. A modern data platform separates storage from computation while colocating them in the same zone and region: data flows from storage such as Cloud Storage or BigQuery to computation such as Dataflow or Managed Service for Apache Spark and back to the target, reducing latency, transfer cost, and idle compute.
Custom builds and managed purchases differ in control, cost, and operational ownership. Custom builds on Compute Engine or GKE, such as Oracle Database on VMs, custom Windows bring-your-own-license images, and custom GKE workloads, give maximum configuration control but require designing availability, replication, failover, and observability from scratch. Managed purchases such as Cloud SQL, Spanner, Cloud Storage, BigQuery, and Dataflow shift that work to Google Cloud with uptime SLAs and built-in backup and resilience. Total cost includes sizing for disk, throughput, IOPS, and vCPUs, licensing where Cloud SQL includes licenses while Oracle on VMs requires customer licenses, and storage type where SSD fits most uses and HDD may fit latency-insensitive datasets above 10 TB. Assessment tools include Google Cloud Migration Center for TCO, migVisor for complexity and phased plans, and Database Migration Assessment for effort estimation. Aggregate availability of chained tiers is tier1_availability x tier2_availability x tierN_availability, so four tiers at 99.9% each yield about 99.6%, while parallel services use 1-(1-A)x(1-B) for higher availability.
Secure archival and ordered teardown remove systems that no longer justify cost, risk, or operational load without breaking dependents. Cloud Monitoring dashboards and Active Assist recommendations flag idle Cloud SQL instances, overprovisioned capacity, unneeded high availability or disaster recovery, unneeded tables, and instances strained by very large table counts. Archival moves the final state to Cloud Storage before deletion, with a final backup of each source database, Bucket Lock retention, holds, versioning, uniform bucket-level access, blocked public access, and open export formats such as CSV, Parquet, and Avro for exit portability. After cutover, applications are monitored for a defined period with fallback planning, and writes can replicate back to the source until retirement is confirmed. Teardown deletes from edge inward in dependency order:
Success measurements turn business goals into trackable technical targets and financial outcomes. A key performance indicator describes the business outcome, a service-level indicator measures system performance, a Service Level Objective sets the desired value over time, and a service-level agreement is the contractual commitment the architecture must support.
Business-aligned indicators and SMART targets link technical work to revenue, cost, risk, and satisfaction. Technical KPIs include latency, throughput, error rates, utilization, and availability, while business KPIs include revenue, costs, conversion, retention, and satisfaction. Cloud Monitoring collects availability, performance, and latency metrics so both types are observed together. System availability protects revenue and trust, transaction throughput supports growth and peaks without over-provisioning, and user engagement signals satisfaction and retention. Objectives are most useful when Specific, Measurable, Achievable, Relevant, and Time-bound, such as 99.9% availability or average response below 200 ms. Quality such as prediction accuracy or F1 score must be separated from performance such as p90 inference latency below 100 ms at peak, so a highly accurate model is not chosen if it is too slow or costly to run.
Reliability indicators, objectives, and budgets measure user-perceived performance and control release risk. The four golden signals are latency, traffic, error rate, and saturation of CPU, memory, GPU, or TPU. An availability SLI is the ratio of successful responses to all responses, while a latency SLI is the ratio of calls below a threshold to all calls. Request-based SLIs count good events over total events, while windows-based SLIs count good time windows over total windows. An SLO in Cloud Monitoring combines the SLI, a performance goal, and a compliance period such as a rolling 30-day period or calendar week, with up to 500 SLOs per service and a documented maximum target of 99.9%. The error budget starts at 1 minus the SLO and declines as performance misses the goal, so teams slow changes when budget runs low and move faster when budget remains. End-to-end monitoring, health checks, load balancer autohealing, and Kubernetes liveness and readiness probes reflect what users experience, with Cloud Logging, Cloud Trace, and Cloud Profiler adding diagnostic context.
Exam tip: An SLI is the measured metric, such as availability percentage or latency, while an SLO is the target for that metric, such as 99.9% or less than 200 milliseconds.
Ownership models, optimization, and validation connect Total Cost of Ownership to return on investment over time. TCO compares full on-premises costs including hardware, software, power, cooling, maintenance, and support with full cloud costs including elastic scaling and network utilization estimated with the price calculator. ROI-focused KPIs separate cost metrics such as cost per inference from business-value metrics such as revenue lift and satisfaction, tracked with Cloud Billing, BigQuery, and Looker under FinOps accountability. Cost visibility organizes resources by organization, folders, and projects with tags such as team=banking_appdev, exports billing to BigQuery, and allocates 100% of costs to spending sources through chargeback, budgets, and regular reviews. Optimization includes autoscaling, rightsizing, custom machine types, Object Lifecycle Management and Autoclass, and discounts where Spot VMs are more than 80% cheaper for fault-tolerant work, resource-based Committed Use Discounts cover Compute Engine quantity in a region, and spend-based discounts cover product spend at the billing account level. Proofs of concept validate hypotheses with uncommon and corner cases, performance and failover requirements, and measurable impact before full deployment, followed by continuous monitoring, data-driven adjustment, iterative refinement, and stakeholder feedback.
Security and compliance integrates data placement, identity governance, and network perimeter controls to meet organizational policies and regulatory requirements. A primary location handles regular traffic while a standby location provides extra capacity or recovery, and placement controls availability, update delay, and whether regulated data stays inside a required boundary.
Placement for bursting, recovery, and sovereignty determines where application copies and database copies run. Cloud bursting scales client traffic across locations when demand exceeds the primary, with application instances in both locations and the public cloud adding elastic capacity to limited on-premises compute. In primary-only deployment the standby application accesses the database in the primary location, while in primary-and-standby deployment each location accesses its local instance kept transactionally synchronized for immediate consistency or asynchronously synchronized with delay and conflict resolution for concurrent writes. A whole cloud is less likely to fail than a single region, but primary plus secondary region plus primary plus secondary cloud lets an enterprise choose the best recovery path for different outage sizes. When data must stay in-country and one cloud has only one region there, bursting into another cloud's region in the same country keeps data inside the country while adding capacity.
Least-privilege identity and sensitive data controls give each person and workload only the permissions its role requires. Cloud Identity authenticates users and can federate with an existing provider, while IAM authorizes actions on each resource with default deny. An administrator may receive full bucket and cluster access while a regular user receives Composer User for DAG and Airflow UI access only, with Airflow UI Access Control further limiting visibility to individual DAGs. Each environment uses its own dedicated service account because DAG code runs with that account's permissions, and Editor or reused default accounts must be avoided. Fine-grained BigQuery access combines table-level IAM with column-level policy tags for Sensitive Personal Identifiable Information, Personal Identifiable Information, Sensitive Personal Information, and Personal Information, where tagged columns remain inaccessible until the policy-tag reader role is granted. A scheduled detector scans catalog tags, publishes findings as JSON to Pub/Sub, loads them via Dataflow to a BigQuery events table, and triggers remediation. Secret Manager (opens in a new tab) stores connections, keys, passwords, and certificates outside DAG files, while bucket, snapshot, and external SSL or TLS paths are limited to trusted users.
Edge protection and isolation boundaries block unwanted traffic, prevent exfiltration, and verify identity before backend access. Traffic flows from internet clients to Google Front Ends to a Google Cloud Armor policy to backend services or buckets, so Layer 3 through Layer 7 screening, DDoS mitigation, WAF, IP and geo filtering, and rate limiting block malicious requests before backend runtimes. VPC firewall rules separately allow or deny traffic to and from backend VMs, including allowing only load balancer and health-check sources to backends. VPC Service Controls creates a service perimeter that allows free communication within the perimeter but by default blocks communication across the perimeter to managed services, controlling internet-to-resource, VM-to-API, and service-to-service paths without requiring a VPC network. A copy between two Cloud Storage buckets succeeds only when both buckets are in the same perimeter. Identity-Aware Proxy verifies identity at backend services and replaces open SSH and RDP with TCP forwarding restricted to IAP ranges. For a global external Application Load Balancer, Cloud Armor evaluates first then IAP, while for a classic Application Load Balancer, IAP evaluates first then Cloud Armor. Deny-by-default egress, skipped default networks, and Private Google Access complete the perimeter.
Continue with this study guide: Security and compliance
Observability uses logs, metrics, traces, and profiles to show health, explain slowness, and trigger response before failures cascade. Cloud Logging (opens in a new tab) stores, searches, and alerts on log events, while Cloud Monitoring collects, stores, visualizes, and alerts on metric data, with telemetry flowing from Compute Engine, GKE, Cloud Run, and App Hub applications to dashboards and alerting policies.
Telemetry collection across runtimes combines logs, metrics, and traces into complementary signals. Metric data measures numeric health at intervals such as CPU and latency, log data records timestamped events with rich detail about one component, and trace data connects spans for single operations into an end-to-end request path for root-cause analysis. Log entries flow from VMs, GKE, Cloud Run, and Cloud Run functions into Cloud Logging for Logs Explorer and custom dashboards, feeding Error Reporting for error groups and log-based metrics for counts. Metric data flows to Cloud Monitoring (opens in a new tab) as system metrics, Ops Agent host and third-party metrics for Apache, Nginx, MongoDB, PostgreSQL, and JVM, user-defined custom metrics via API or OpenTelemetry, external metrics, Prometheus metrics via Managed Service for Prometheus, and log-based metrics. The Ops Agent extends Compute Engine with host, log, Prometheus, and OTLP data, while Cloud Run uses Prometheus or OpenTelemetry sidecars and GKE uses observability packages for control-plane and SLO signals.
Exam tip: Choose user-defined custom metrics when application code must emit a new numeric signal through the API or OpenTelemetry, and choose log-based metrics when the signal can be counted or measured from log entries already written to Cloud Logging.
Objectives, alerts, and incident records turn measurements into operational controls. An availability SLI divides successful responses by all responses, while a latency SLI divides calls below the threshold by all calls, with request-based SLIs counting atomic units and windows-based SLIs counting good time periods. An SLO adds a performance goal and compliance period, such as latency exceeding 300 ms in only 5 percent of requests over 30 days, with up to 500 SLOs per service and a maximum target of 99.9%. The error budget starts at 1 minus the SLO, so a 99.9% SLO allows 0.1% imperfect performance before the period fails. An alerting policy defines metric or burn-rate criteria and notification recipients, sending through email, Cloud Mobile App, PagerDuty, or Slack and opening an incident with status, duration, labels, metric charts, and log links. Policies can be managed in console, API, CLI, or Terraform, with preconfigured recommended alerts for many services.
Latency, efficiency, and failure diagnostics explain why distributed systems are slow or failing. Cloud Trace follows a single request from entry-point through downstream microservices in near real time, recording time in each span to pinpoint the slow service, with shared trace and span IDs linking traces to logs. Cloud Profiler continuously profiles function blocks to show which functions consume the most CPU or memory, connecting costly code to saturation of CPU, GPU, TPU, and memory. Error Reporting aggregates repeated crashes from GKE, Cloud Run, and App Engine into error groups with counts, history, and stack traces linked to logs. Together with the four golden signals of latency, traffic, error rate, and saturation, plus CPU, GPU, TPU, memory, network, and disk tracking, these tools distinguish slow requests between services from inefficient code within a service and recurring failures that need root-cause fixes.
Continue with this study guide: Observability
A process flow showing a user request going through Retrieve, Augment, and Generate steps, with SAP function modules and Vector Search feeding facts into Retrieve and Gemini models supporting generation.
A service level indicator (SLI) measures behavior from the user perspective such as percentile latency or successful-response ratio, while a Service Level Objective (SLO) sets the target with a bound and time window. For example, an SLO can require 95th percentile latency below 300 ms over a rolling 30-day period, and Cloud Monitoring supports up to 500 SLOs per service with a documented maximum target of 99.9 percent.
Use SAP function modules when the dataset is small and resides within SAP, and use Vector Search products when enterprise data is large and low latency is required. With function modules data flows from SAP tables through SELECT queries, BAPI calls, or Gemini function calling to the serving subsystem, while with Vector Search embeddings flow to a vector database such as Cloud Storage, Agent Platform Feature Store, or BigQuery and semantic search results flow to the serving subsystem.
Cold standby keeps only backups in Cloud Storage, images, and templates with no running servers and has the largest Recovery Time Objective (RTO) and Recovery Point Objective (RPO) at the lowest cost. Warm standby runs a minimal database that receives replication with app servers restored from snapshots for small RTO and RPO at medium cost, while hot standby runs a full secondary environment that serves traffic with continuous replication for near-zero RTO and RPO at the highest cost.
Use Storage Transfer Service for fully managed online copies optimized above 1 TiB and able to scale to petabytes, and use Transfer Appliance for offline moves recommended above 20 TB up to 1 petabyte. Storage Transfer Service depends on allocated bandwidth with parallelization, retries, and integrity validation, while Transfer Appliance is loaded locally and shipped to Google to avoid bulk network use and business disruption.
Last updated