Prepare and test your skills
Prepare and test your skills
Application and infrastructure deployment is the automated, repeatable delivery of application code and cloud infrastructure across Virtual Private Clouds (VPCs), regions, and clouds. Code flows from development through a continuous integration and continuous delivery (CI/CD) pipeline into regional application stacks, while infrastructure flows from declarative files into consistent environments. The regional stacks, access VPCs, and framework deployments below work together so every release lands on identical, resilient infrastructure.
Shared and dedicated service access controls how broadly a managed service can be reached from VPC networks. A service access VPC (opens in a new tab) hosts shared consumer access points with transitive connectivity to other VPCs and external networks, while an application VPC hosts dedicated access points reachable only inside that same VPC. A VPC that hosts both application resources and shared access points must be treated as a service access VPC because transitive routing changes its firewall responsibilities.
| Access mode | Deployment location | Reachability |
|---|---|---|
| Shared consumer access points | Service access VPC | Accessible from any other VPC or external network |
| Dedicated consumer access points | Application VPC | Accessible only from resources within that application VPC |
Traffic flows from application VPCs and external networks to forwarding rules in the services VPC for shared access, so shared deployments depend on correct route advertisements and firewall rules. For high availability virtual private network (HA VPN) connections, the Cloud Router advertises the consumer access subnet as a custom route, and multicloud firewalls must allow the private services access subnet. For serverless access, the VPC Access connector requires a dedicated /28 subnet, and ingress rules must open that connector subnet within the VPCs and toward the application VPCs. Dedicated access is simpler because traffic flows only inside one application VPC to its local forwarding rule, but the rule and firewall entries must be repeated in every VPC that needs the service.
A regional application stack is a concatenation of service layers that all operate in the same region, such as a web layer calling an application layer calling a database layer in that region. Each layer uses load balancers to spread traffic across endpoints in three or more zones, and each service uniform resource locator (URL) resolves through Domain Name System (DNS) to the regional IP address of its layer load balancer. The same stack pattern is repeated in each region across Google Cloud, other clouds, and on-premises data centers.
This repetition enables active-standby resiliency across regions. Out-of-band monitoring watches every layer, and when it reports a failure in any layer, the application frontend is re-anchored to the complete backup stack in another region. Automated deployment must therefore keep regional stacks identical, or failover will behave differently from normal operation.
Infrastructure as Code (IaC) defines cloud resources in versioned files and provisions them with tools such as Terraform or Deployment Manager. In Google Cloud, a framework deployment applies a versioned compliance framework and its cloud controls to an organization, folder, or project using the gcloud compliance-manager framework-deployments create command, a stored metadata file, and explicit target flags. The same definition can be reused for development, staging, and production to prevent drift.
The metadata file separates desired state from the deployment command. Each control entry names a control revision and enforcement behavior:
name: full resource name such as organizations/ORGANIZATION_ID/locations/LOCATION/cloudControls/CLOUD_CONTROL_NAMEmajorRevisionId: major version such as 1 for restrict-bucket-regionparameters: optional settings such as a location string value of us-westenforcementMode: AUDIT, DETECTIVE, or PREVENTIVEThe deployment object lives in an organization and location and enforces controls on the target selected by flags. Use --target-resource-config-existing for an existing organization, folder, or project, use folder-creation flags with display name and parent to create a folder during deployment, and use project-creation flags with display name, parent, and billing account to create a project during deployment.
Exam tip: Shared access points live in a service access VPC and are reachable broadly, while dedicated access points live in an application VPC and are reachable only inside that VPC.
Progressive delivery moves new versions to users in small steps so problems are found before full rollout. Cloud Deploy automates this with a delivery pipeline that holds canary configuration, a release that represents the new container image, and a rollout that moves the release through ordered phases such as canary-50 followed by stable. Traffic flows first to the small canary population and only after approval flows to the stable population. Cloud Deploy skips the canary phase and runs only stable on the very first deployment to a runtime because there is no prior version for comparison or fallback.
Load balancing protects availability during those rollouts. Traffic flows from the load balancer to backends such as instance groups, Google Kubernetes Engine (GKE) pods, or managed endpoints, and the load balancer depends on health checks to decide where traffic can go. When a check marks a backend unhealthy, the load balancer stops sending it traffic and sends traffic only to healthy resources. Services such as App Engine and Cloud Run balance automatically, while Compute Engine and GKE require selection of a load balancer by traffic type. Multiple environments — production, staging, testing, development, and canary — plus GKE release channels and regional or multi-regional redundancy further isolate validation from live traffic and survive zone or region loss.
Continue with this study guide: Application and infrastructure deployment
API management best practices use Apigee as a proxy layer between client applications and backend services to enforce security, control traffic, and keep public interfaces stable. Traffic flows from clients to the Apigee proxy and then from the proxy to backends, so producers can change implementation without breaking consumers. Policies for authentication, quotas, caching, and mediation execute at the proxy, which protects backends and simplifies operations.
Security and traffic controls at the proxy decide who can call an API and how much load can reach backends. API key validation checks the request.queryparam.apikey variable against an API product that includes the proxy; only requests with a valid key from a registered developer flow to the backend, while unresolvable keys return a fault such as steps.oauth.v2.FailedToResolveAPIKey. Rate limiting caps API requests over time with quotas, which protects services from malicious distributed denial-of-service (DDoS) traffic and from spikes caused by client configuration errors, and project-level quotas also limit billing exposure.
Throttling supports graceful degradation, where a system under load keeps working with reduced performance instead of failing completely. Each Apigee replica independently slows incoming requests under policy control, excess requests are dropped early at the frontend to protect backends, partial errors and retries are handled seamlessly, overload is simulated regularly, and API monitoring (opens in a new tab) tracks spikes before they escalate. Dropping some requests early limits impact to errors for some users, in contrast to circuit-breaking where all traffic is dropped during overload.
A stable interface isolates consumers from backend evolution through the API proxy (opens in a new tab) as a mapping from a public HTTP endpoint to a backend service. Clients need only the proxy URL, parameters, credentials, and response format, so backend moves such as changing hosts, databases, or code do not change the client contract. The same proxy layer supports REST, gRPC, SOAP, and GraphQL styles, letting producers serve many platforms while operations applies consistent policy.
The developer portal provides self-service onboarding for that stable interface. Client developers discover published APIs, read documentation needed to build apps, and register to stay in sync with updates without direct coordination with backend teams. Operational visibility closes the loop because all runtime traffic passes through containerized services that emit logging, metrics, and analytics for uptime, usage, and quota decisions.
Mediation decouples client-facing APIs from backend speed and capacity through the Apigee gateway. A request follows an ordered path:
Because the response must traverse the gateway in both directions, a slow backend directly increases end-to-end latency. An HTTP(S) Load Balancer in front of the Apigee Istio Ingress Gateway handles protocol compatibility at the edge by accepting HTTP/1.0 from older clients and forwarding as HTTP/1.1 or HTTP/2 to Apigee. Response caching further reduces load by serving repeated reads directly from the gateway without calling the backend.
| Backend problem | Cause | Impact on API traffic |
|---|---|---|
| Inadequately sized backend | Backend lacks CPU, memory, or capacity for API demand | Traffic spikes lead to failed requests |
| Slow backend | Backend is improperly tuned and responds very slowly | Latency increases with timeouts and degraded experience |
Routing scope controls failover and residency. Global routing uses Layer 7 global external load balancing with automatic failover across regions, while Regional routing restricts balancing to the Apigee instance region for data residency. Network Service Tiers Premium Tier carries internet traffic on the Google backbone from the point of presence (PoP) closest to the sender to the PoP closest to the destination, reducing hops for global users, while internal load balancers keep web-to-app traffic in a 3-tier stack off the internet.
Exam tip: Global routing is selected by default unless advanced data residency is enabled, in which case Regional routing is selected by default and cannot be changed.
Continue with this study guide: API management best practices (e.g., Apigee)
Testing frameworks validate reliability, performance, and scalability by simulating real load and by checking every change before promotion. Distributed load testing proves that autoscaling, health checking, rate limiting, and network paths meet Service Level Objectives (SLOs) under stress. Unit, integration, and end-to-end tests in CI/CD pipelines catch logic and connection errors before they reach production.
Validating autoscaling means measuring successful responses and latency, not just infrastructure size. Autoscaling adds resources for current load, but an application can still be unresponsive when virtual machines (VMs) are in the RUNNING state while application software on some VMs has crashed. Cloud Load Balancing (opens in a new tab) health checks ensure traffic flows only to responsive backends, and for managed instance groups (MIGs) autohealing proactively deletes unavailable VMs and creates replacements. A valid load test therefore exercises scaling, health-check failures, and autohealing together against SLOs.
Distributing stress across tiers is equally important. Load balancing improves reliability by routing only to available resources and preventing individual resources from overloading. In a 3-tier stack, an internal load balancer carries web-to-app traffic, so tests must stress internal paths as well as external client paths. In a global topology, traffic flows from internet clients to web servers to application servers to a synchronously replicated database such as Spanner, with global external balancing to web servers by availability, capacity, and proximity and cross-regional internal balancing to app servers by availability and capacity.
The network path directly affects benchmarked latency and must match production. In Premium Tier, internet traffic enters the Google network at the PoP closest to the sender, travels within Google to the Compute Engine resource, and exits at the PoP closest to the destination, reducing hops and improving perceived availability. Standard Tier uses the public internet at lower cost. Benchmarks run on the wrong tier are not comparable to production SLOs.
Matching the load balancer to the workload means choosing by traffic type, client location, and backend placement. Application Load Balancers are proxy-based Layer 7 balancers for HTTP(S) behind a single anycast IP, while Network Load Balancers are Layer 4 balancers for TCP, UDP, and other IP protocols with proxy variants that terminate traffic and passthrough variants that preserve client IP and support direct server return. Narrow the choice by external versus internal scope, global multi-region versus regional single-region backends, Premium versus Standard Tier, and needs such as TLS termination location, performance, and cost.
Exam tip: A global load balancer supports backends in multiple regions, while a regional load balancer supports backends in a single region but remains globally accessible.
Environment parity makes test results valid by keeping test and production configurations identical. Infrastructure as Code with tools such as Terraform defines networks and baselines in code, so dedicated test environments can be created on demand and torn down after integration tests without manual drift. Cloud Build executes builds and automated unit, integration, and end-to-end tests on each change, and Cloud Deploy promotes only validated artifacts through ordered target environments. Tool selection affects parity through stack compatibility, automation degree, CI/CD integration, and defect and program management.
Test data must be realistic yet protected. Post-migration checks run pipelines with production-grade input and confirm completeness, such as time-series data no more than five minutes behind current time, while Sensitive Data Protection classifies financial and privacy-regulated data for GDPR and CCPA. Data-layer validation runs aggregations in BigQuery or Dataflow in both source and target and compares results, using a client that pulls from both regions when tables cannot be joined directly. Managed engines such as BigQuery and Dataflow need mainly result validation, while non-managed engines such as Spark, Hive, Flink, and MapReduce need per-runtime testing. Local emulators for Bigtable, Datastore, Firestore, Pub/Sub, and Spanner enable isolated integration tests without service cost or production impact.
Continue with this study guide: Testing frameworks (load/unit/integration)
Data and system migration tooling moves files, warehouses, servers, and delivery processes to Google Cloud with controlled downtime and rollback options. Large-scale data flows from source staging to Cloud Storage and then from Cloud Storage to BigQuery, where schema and data become queryable tables. Infrastructure and application delivery flows from Infrastructure as Code foundations through build, registry, and promotion services to production runtimes.
Online versus offline transfer is the choice between network delivery and shipping physical hardware. Online transfer uses the internet, Direct Peering, Cloud Interconnect, or cloud-to-cloud links, while offline transfer fills Transfer Appliance hardware on site and ships it to a Google upload facility. The decision depends on dataset size, bandwidth, deadline, and cost: 100 TB takes over 10 days at 1 Gbps and over 100 days at 100 Mbps, so constrained bandwidth can make offline faster despite shipping delay.
| Source and condition | Online or offline | Choose |
|---|---|---|
| Another cloud to Google Cloud | Online | Storage Transfer Service |
| Cloud Storage bucket to another bucket | Online | Storage Transfer Service |
| Private data center with enough bandwidth | Online | gcloud storage for small to medium, Storage Transfer Service for large scale |
| Private data center without enough bandwidth | Offline | Transfer Appliance |
The staging flow preserves schema and integrity in order: first extract schema and data to vendor-neutral files, then upload files to a Cloud Storage bucket closest to the source, then load from the bucket to BigQuery by batch load or scheduled transfer. Avro, ORC, and Parquet embed schema with data and reduce complexity, while CSV requires separate schema or auto-detection. The gcloud storage command suits interactive, program-driven, and low-latency sync work with parallel uploads and resume, while Storage Transfer Service suits managed, scheduled, petabyte-scale moves with agents on common NFS mounts, bandwidth caps, error reports, and TSV records. BigQuery Data Transfer Service (BigQuery DTS) then loads into BigQuery tables directly or on schedule, including from Redshift and Teradata with Cloud Storage staging.
Exam tip: Use Storage Transfer Service to move objects into Cloud Storage, use BigQuery Data Transfer Service to load into BigQuery tables, and use Transfer Appliance only when bandwidth cannot meet the deadline.
Provisioning migrated foundations with Infrastructure as Code defines hierarchy, networking, identity, and security in versioned files instead of console steps. Teams choose a starting point from provider resources for maximum control to blueprints and Enterprise Foundations for opinionated speed. Once chosen, foundations are built in order: first build organization, folder, and project hierarchy, then configure Identity and Access Management (IAM), then set up billing, then establish source-to-Google connectivity, then harden security, then enable logging, monitoring, and alerting.
Automating delivery moves artifacts from source to runtimes without manual handoffs. Cloud Build imports source and produces containers or archives, Artifact Registry stores packages and images as the single location integrated with CI/CD and runtimes, and Cloud Deploy promotes releases to GKE, Cloud Run, App Engine flexible environment, and Cloud Run functions. The build depends on the repository, deployment depends on the registry artifact, and the runtime depends on the promoted release. Artifact-first refactoring preserves rollback by first provisioning Artifact Registry repositories, then dual-storing artifacts in source and Google registries, then deploying a small subset to Google and expanding, then storing only in Artifact Registry, then migrating older versions with tools such as gcrane, then decommissioning source repositories.
Migration strategy matches each workload to downtime tolerance, tool fit, and cutover order. Mission-critical workloads that cannot tolerate downtime need zero or near-zero downtime strategies that sync in the background, while tolerant secondary workloads can use simpler downtime-window cutover. Zero-downtime designs reduce disruption but cost more and add complexity. Tool choice weighs strategy, source and target environments, data size, change frequency, and managed-service availability.
| Tool | Strategy | Source to destination |
|---|---|---|
| Migrate to Virtual Machines (opens in a new tab) | Lift and shift | On-premises VMware and other-cloud VMs to Compute Engine |
| VMware HCX | Lift and shift | VMware VMs to Google Cloud VMware Engine as continuous hybrid infrastructure |
| Migrate to Containers | Replatform | Physical servers and VMs to GKE, Autopilot, GKE Enterprise, and Cloud Run |
Execution follows assess, plan, and repeatable deploy-test-optimize cycles. Assessment inventories workloads, tests proofs of concept, and calculates total cost including network charges and rightsizing savings. Planning divides work into waves of related applications and move groups that migrate together in sprints with batch order, design documents, and timelines. Each sprint assesses dependencies, plans checklist and rollback with a maximum time trigger, deploys and tests with gradual rollouts, and optimizes by capturing lessons for the next wave.
Continue with this study guide: Data and system migration and management tooling
Gemini Cloud Assist (opens in a new tab) is an AI-powered collaborator that helps design, troubleshoot, secure, and optimize Google Cloud solutions. Diagnostic and cost data flows from Google Cloud resources and APIs into Gemini Cloud Assist, and guidance flows back as answers, hypotheses, and recommended fixes. It reasons with live environment context, remembers conversation history, and acts with the same permissions as the requesting user.
Investigations are AI-assisted troubleshooting sessions that compare current logs, metrics, and performance against baseline behavior. An investigation in the Cloud Assist panel (opens in a new tab) depends on access and context: the user asks for help, Gemini Cloud Assist offers a Create investigation button when useful, the user confirms pre-populated resources and time range, and analysis runs to produce observations and hypotheses that can be attached to a support case. Chat-based investigation requires a Premium Support contract or approved access plus the Investigation Creator role.
Bigtable investigations target cluster performance for a selected cluster and time range. The operator opens the instance, opens System insights, selects Explore Investigations, and starts analysis of client latencies, server latencies, CPU utilization, disk utilization, and request count. Results separate observations such as channel pool misconfiguration from Hypotheses with recommended actions. Managed Service for Apache Spark investigations, formerly Dataproc on Compute Engine and Serverless for Apache Spark, target failed and slow batches using logs, metrics, and operator context. The INVESTIGATE button appears for Failed batches on the Batches List page and for Succeeded and Failed batches in the Batch Details Health overview, changing to INVESTIGATING during analysis and VIEW INVESTIGATION when ready. Results organize into Issue, Relevant Observations, and expanded Hypotheses, and the operator applies fixes and reruns the job.
| Investigation type | What it analyzes | Where it starts |
|---|---|---|
| Bigtable investigation | Cluster performance over a time range using latency, CPU, disk, and request signals | Instance System insights and Explore Investigations |
| Managed Service for Apache Spark investigation | Failed or slow batch using logs, metrics, and context | Batches List or Batch Details Health overview |
Security posture work with Gemini Cloud Assist finds weaknesses, explains risks, and guides remediation under the requester's own permissions. Page context awareness uses the current console page URL and visible text, while memory preserves session details across turns, so follow-ups stay tied to the same cluster, image, or policy. Because visibility is permission-bound, incomplete findings often mean missing roles such as Cloud Asset Viewer for topology via Cloud Asset Inventory or Security Auditor for posture data.
Artifact review queries Artifact Analysis scan results for repositories in Artifact Registry to answer which images contain a vulnerability or package and to show build provenance and software bill of materials (SBOMs). Deeper misconfiguration analysis runs as an investigation that starts from chat, Logs Explorer, an error, or a resource, then analyzes logs and configurations, ranks hypotheses, and recommends next steps transferable to a support case. Guided remediation then generates IAM role advice, CLI commands, kubectl manifests, and Terraform code with in-chat simulation, but every suggestion must be validated before use because plausible output can be incorrect.
| Role | Can invoke and chat | Can manage artifacts and configuration |
|---|---|---|
| Gemini Cloud Assist Viewer | No, view only | No |
| Gemini Cloud Assist User | Yes, invoke and share owned artifacts | No |
| Gemini Cloud Assist Editor | Yes, with User permissions | Yes, delete topics and update agent configs |
| Gemini Cloud Assist Admin | Yes, with Editor permissions | Yes, plus policies and project-wide sharing |
Exam tip: Gemini Cloud Assist sees only what the requester can see, so add Cloud Asset Viewer plus the product-specific role when results are incomplete.
Architectural optimization connects design questions to live inventory, topology, and spend data. A natural-language question about cost flows from chat to cost sources, and the answer links products, SKUs, resources, and workloads to specific actions. Features include chat-based cost optimization, FinOps Hub views with AI summaries, daily proactive anomaly analysis that identifies the spike owner and fix, database rightsizing insights, flow analysis in Network Intelligence Center, and Storage Insights for bucket and object usage.
Resource-aware recommendations depend on foundation APIs that supply inventory and cost context. The Cloud Asset API provides resource and topology inventory, the Design Center API enables configuration generation, the App Optimize API provides cost-utilization data, and the App Hub API scopes queries to applications and workloads. The Cloud Monitoring API, Cloud Logging API, Recommender API, and App Topology API add metrics, audit logs, usage recommendations, and topology detail. Without these APIs, answers are more limited. Investigations complement chat with ordered initiation, data gathering, hypothesis ranking, recommendation, and support handoff, powered by agentic loops that iteratively call Cloud and Kubernetes APIs while preserving multiturn context.
Continue with this study guide: Gemini Cloud Assist
Shared consumer access points live in a service access Virtual Private Cloud (VPC) and are accessible from any other VPC or external network, while dedicated consumer access points live in an application VPC and are accessible only from resources within that same application VPC. A VPC that hosts both application resources and shared access points must be treated as a service access VPC because transitive routing changes its firewall responsibilities.
Cloud Deploy moves a release representing a new container image through ordered phases such as canary-50 followed by stable, sending traffic first to the small canary population and only after approval to the stable population. Cloud Deploy skips the canary phase and runs only stable on the very first deployment to a runtime because there is no prior version for comparison or fallback.
Use Storage Transfer Service to move objects into Cloud Storage, including online moves from another cloud, between buckets, or large-scale moves from a private data center with enough bandwidth, and use BigQuery Data Transfer Service (BigQuery DTS) to load into BigQuery tables directly or on schedule. Use Transfer Appliance only when bandwidth cannot meet the deadline, such as a private data center without enough bandwidth where 100 terabytes would take over 10 days at 1 gigabit per second.
Global routing uses Layer 7 global external load balancing with automatic failover across regions, while Regional routing restricts balancing to the Apigee instance region for data residency. Global routing is selected by default unless advanced data residency is enabled, in which case Regional routing is selected by default and cannot be changed.
Last updated