PCA — Google Cloud Professional Cloud Architect 2026 Cheat Sheet

Cheat sheet: Google Cloud Professional Cloud Architect 2026 (PCA) architecture reference for service selection, security, reliability, networking, and exam decision points.

Use the tables for a quick pre-exam check. Expand a topic’s notes for explanations, examples, and additional distinctions.

Scope and study context

For PCA-style questions, read for:

  1. Business goal: reduce operations, improve availability, meet compliance, cut cost, modernize.
  2. Workload shape: stateless, stateful, batch, streaming, OLTP, analytics, ML, legacy.
  3. Constraints: latency, RTO/RPO, data residency, existing tooling, skills, portability.
  4. Managed-service fit: prefer managed/serverless when requirements allow.
  5. Security and governance: least privilege, centralized control, auditability, network boundaries.
  6. Tradeoff: cheapest is not always correct; most managed is not always correct.

The PCA exam is not just a service-name recall test. Strong candidates can translate business requirements into secure, reliable, cost-aware Google Cloud architectures. Expect scenarios where several answers are technically possible, but only one best matches the constraints.

  • Core Google Cloud architecture decision rules
  • Compute, networking, data, security, and operations patterns
  • Common traps in scenario questions
  • How to connect review to IT Mastery practice and original practice questions

Core PCA Architecture Heuristics

If the scenario emphasizes…Prefer…Watch for traps
Minimal operationsCloud Run, App Engine, Cloud SQL, BigQuery, Pub/Sub, managed servicesDo not choose GKE/Compute Engine just because they are familiar
Existing VM app with OS dependenciesCompute Engine, managed instance groups, Migrate to Virtual MachinesDo not force serverless if custom kernel, agents, or stateful local dependencies are required
Container portability and Kubernetes ecosystemGKE Autopilot or GKE StandardAutopilot reduces ops; Standard gives more node-level control
Simple event-driven functionCloud Run functions / Cloud FunctionsFor container control, concurrency, or portable runtime, use Cloud Run
Global relational consistency and scaleCloud SpannerCloud SQL is simpler when a single-region relational database is enough
Analytics over large datasetsBigQueryBigQuery is not an OLTP database
Streaming ingestionPub/Sub plus DataflowPub/Sub alone is messaging, not transformation/analytics
Low-latency key/value or wide-column accessBigtableNot relational; not ad hoc SQL analytics
Centralized enterprise networkShared VPC, Cloud Interconnect/VPN, Cloud Router, Network Connectivity CenterVPC peering is not transitive
Data exfiltration protectionVPC Service Controls, IAM, org policy, loggingVPC Service Controls do not replace IAM
Hybrid private connectivityCloud Interconnect or HA VPNCloud NAT is outbound only; it does not provide inbound access
Global web appGlobal external Application Load Balancer, Cloud CDN, Cloud ArmorRegional LBs may be wrong for worldwide latency and failover
Cost control for batch/fault-tolerant computeSpot VMs, autoscaling, managed services, lifecycle policiesSpot VMs can be interrupted; design for retry

Service Selection: Compute

ServiceChoose whenAvoid whenPCA cues
Cloud RunStateless containerized services, APIs, webhooks, event consumers, scale-to-zero, low opsStateful workloads, deep VM control, specialized networking not supported by service“Containerized app,” “serverless,” “variable traffic,” “reduce operations”
Cloud Run JobsContainerized batch jobs, scheduled tasks, finite workLong-running services needing continuous serving“Run a job,” “batch container,” “no server management”
Cloud Run functions / Cloud FunctionsLightweight event handlers, glue code, simple HTTP functionsComplex services, custom container needs, extensive runtime control“Trigger on upload/message,” “small function”
App Engine StandardWeb apps with supported runtimes, automatic scaling, minimal infrastructureCustom OS/packages, unsupported runtime, portability concerns“Legacy App Engine,” “simple web app,” “PaaS”
App Engine FlexibleApp Engine model with custom runtime/container-like flexibilityNeed Kubernetes control or scale-to-zero economics“Managed web app but custom runtime”
GKE AutopilotKubernetes workloads with reduced node managementNeed detailed node, daemonset, privileged, or specialized control“Kubernetes required,” “reduce cluster operations”
GKE StandardFull Kubernetes control, custom node pools, GPUs, service mesh, complex networkingTeam lacks Kubernetes skills; serverless fits better“Kubernetes ecosystem,” “custom scheduling,” “platform team”
Compute EngineLift-and-shift, OS-level control, legacy apps, custom agents, specialized machinesApp can be easily containerized/serverless“VM-based,” “custom OS,” “existing licenses,” “stateful server”
Managed Instance GroupsAutoscaled/stateless VM fleets, rolling updates, self-healingSingle pets, unmanaged stateful VMs“Autoscale VMs,” “self-healing,” “regional availability”
BatchLarge batch processing, parallel jobs, queue-based computeAlways-on web serving“Batch workloads,” “parallel compute,” “jobs”
VMware EngineVMware migration with minimal re-platformingCloud-native modernization is feasible and required“Existing VMware,” “rapid data center exit”
Notes and examples

Compute Decision Path

    flowchart TD
	    A[Workload to host] --> B{Needs OS or VM-level control?}
	    B -->|Yes| C[Compute Engine or MIG]
	    B -->|No| D{Requires Kubernetes APIs/ecosystem?}
	    D -->|Yes| E{Need node-level control?}
	    E -->|Yes| F[GKE Standard]
	    E -->|No| G[GKE Autopilot]
	    D -->|No| H{Containerized service or job?}
	    H -->|Service/API| I[Cloud Run]
	    H -->|Finite job| J[Cloud Run Jobs or Batch]
	    H -->|No, simple function| K[Cloud Run functions / Cloud Functions]

Service Selection: Storage and Databases

NeedBest-fit servicesKey distinction
Object storageCloud StorageDurable object store for files, backups, static assets, data lakes
VM block storagePersistent Disk / HyperdiskAttached to Compute Engine or GKE nodes; not object storage
Shared POSIX/NFS file systemFilestoreManaged NFS for apps requiring shared file semantics
Relational database, moderate scaleCloud SQLManaged MySQL, PostgreSQL, SQL Server; simpler than Spanner
PostgreSQL-compatible high performanceAlloyDBManaged PostgreSQL-compatible option for demanding workloads
Global relational scale and strong consistencyCloud SpannerDistributed relational database; choose when Cloud SQL cannot meet scale/availability/global needs
Document databaseFirestoreServerless document model, mobile/web backends, flexible schema
Wide-column, low-latency high throughputBigtableTime series, IoT, personalization, operational analytics; not relational
In-memory cacheMemorystoreRedis/Valkey-compatible cache layer; not durable primary system of record
Data warehouseBigQueryServerless analytics, SQL over large datasets; not OLTP
SearchVertex AI Search / search-oriented managed optionsUse when full-text/search relevance is central, not simple primary-key lookup
Notes and examples

Database Decision Table

RequirementPreferWhy
Existing MySQL/PostgreSQL/SQL Server appCloud SQLLowest migration complexity
Existing PostgreSQL app needs higher performance/scalingAlloyDBPostgreSQL compatibility with managed performance features
Multi-region relational writes and strong consistencySpannerDesigned for distributed relational scale
Mobile app with flexible JSON-like dataFirestoreServerless document database with client-friendly model
Time-series telemetry at high write rateBigtableWide-column, low-latency writes/reads
BI dashboards over petabyte-scale dataBigQueryAnalytical SQL engine
Cache frequent readsMemorystoreReduces database load; not authoritative storage
Lift-and-shift database with full admin controlCompute Engine database VMOnly when managed database does not meet requirements

Cloud Storage Selection

RequirementUseNotes
Frequently accessed objectsStandard storage classCommon for active content and serving
Infrequent accessNearline / Coldline / ArchiveChoose based on access pattern and retrieval expectations; avoid memorizing invented thresholds
Automatic tiering by ageLifecycle rulesCommon exam answer for cost optimization
Prevent accidental deletionVersioning, retention policy, soft delete controls where applicableMatch to governance need
Static web assetsCloud Storage plus Cloud CDN / load balancing where neededBucket alone is not full application hosting
Private object accessIAM, signed URLs, signed policy documentsAvoid making buckets public unless explicitly required
Data lakeCloud Storage plus Dataproc, Dataflow, BigQuery, DataplexObject store is storage; governance/query need additional services

Storage and database decision table

Most PCA data questions come down to workload shape: object storage, relational transactions, global consistency, document access, time-series/wide-column data, or analytics.

WorkloadStrong candidate service
Object storage, backups, media, data lake landing zoneCloud Storage
Managed MySQL/PostgreSQL/SQL Server relational databaseCloud SQL
High-performance PostgreSQL-compatible transactional workloadAlloyDB
Global relational database with strong consistency and horizontal scaleSpanner
Document database for app developmentFirestore
Wide-column, low-latency, high-throughput workloadsBigtable
In-memory cacheMemorystore
Enterprise file sharesFilestore
Analytics warehouseBigQuery
Streaming messagingPub/Sub
Stream/batch data processingDataflow
Managed Spark/HadoopDataproc
Workflow orchestrationCloud Composer
Data lake storage foundationCloud Storage plus analytics tools

Data decision traps

TrapBetter thinking
Choosing BigQuery for high-frequency OLTP app transactionsBigQuery is analytics-focused; use transactional databases for OLTP
Choosing Cloud SQL for global horizontal scale with strong consistencyConsider Spanner if global relational scale is required
Choosing Bigtable because data is “big”Bigtable fits low-latency key-based access patterns, not general SQL analytics
Choosing Firestore for complex relational joinsFirestore is document-oriented; model access patterns carefully
Storing secrets in Cloud Storage or source codeUse Secret Manager
Ignoring data residency or compliance cluesChoose region/multi-region deliberately and apply governance controls
Overusing multi-region storage without needBalance availability, latency, compliance, and cost

Analytics, Data Engineering, and AI

Use caseChooseAvoid confusing with
Decoupled event ingestionPub/SubCloud Tasks for task dispatch with explicit retry/target semantics
Stream/batch transformationDataflowPub/Sub stores messages; Dataflow processes them
Hadoop/Spark migrationDataprocBigQuery for serverless SQL analytics
Serverless warehouseBigQueryCloud SQL for OLTP
Workflow orchestrationCloud ComposerDataflow for data processing; Workflows for simpler service orchestration
Visual/code-light ETLCloud Data FusionDataflow for custom Apache Beam pipelines
CDC into analyticsDatastreamDatabase Migration Service for database migration workflows
SQL data transformation in BigQueryDataformComposer when cross-system orchestration is required
Data governance and discoveryDataplex, Data Catalog capabilitiesIAM alone is not data governance
BI semantic layer/reportingLooker / Looker StudioBigQuery stores and queries data
ML model training/servingVertex AIBigQuery ML for SQL-centric models inside BigQuery
Prebuilt document extractionDocument AICustom ML when prebuilt processors do not fit
Sensitive data discovery/maskingSensitive Data ProtectionKMS encrypts; it does not classify data
Notes and examples

Common Data Architecture Patterns

PatternReference architecture
Real-time analyticsProducers → Pub/Sub → Dataflow → BigQuery → Looker
Batch data lakeSources → Cloud Storage → Dataproc/Dataflow/BigQuery external or native tables → Dataplex governance
CDC replicationSource DB → Datastream → Cloud Storage/BigQuery or target pipeline
ML pipelineCloud Storage/BigQuery → Vertex AI training/pipelines → model registry → endpoint/batch prediction
Log analyticsCloud Logging sinks → BigQuery/Cloud Storage/Pub/Sub → analysis or archival

Networking and Connectivity

Service / featureChoose whenKey exam distinction
VPCPrivate network foundationVPC is global; subnets are regional
Shared VPCCentral networking team, multiple service projectsHost project owns network; service projects deploy workloads
VPC Network PeeringPrivate RFC1918 connectivity between VPCsNot transitive; no central inspection by default
Private Service ConnectPrivate access to Google APIs, producer services, or supported SaaSBetter for controlled private service exposure than broad peering
Private Google AccessVM without external IP needs Google APIs over internal pathApplies to subnet/VM access patterns; not general inbound connectivity
Cloud NATPrivate VMs need outbound internetNo inbound connections; pair with no external IPs
Cloud RouterDynamic routing with BGPUsed with HA VPN and Interconnect
HA VPNEncrypted hybrid connectivity over internetFaster to set up; bandwidth/latency less predictable than Interconnect
Dedicated InterconnectHigh-throughput private physical connectionRequires colocation/provider planning
Partner InterconnectPrivate connectivity through supported providerUseful when Dedicated Interconnect is not practical
Cross-Cloud InterconnectPrivate connectivity to supported other cloudsUse for multicloud private networking scenarios
Network Connectivity CenterHub-and-spoke connectivity managementGood for enterprise hybrid/multicloud topology
Cloud DNSManaged DNSCombine with private zones for internal resolution
Cloud CDNCache static/cacheable content at edgeRequires appropriate backend/load balancing pattern
Cloud ArmorWAF and DDoS protection for supported HTTP(S) workloadsNot a substitute for IAM or app auth
Notes and examples

Load Balancer Selection

RequirementPreferNotes
Global HTTP(S) web appGlobal external Application Load BalancerL7 routing, TLS, Cloud CDN, Cloud Armor integration
Regional HTTP(S) appRegional external Application Load BalancerUse when regional scope is required
Private HTTP(S) servicesInternal Application Load BalancerInternal L7 traffic management
TCP/UDP pass-throughNetwork Load Balancer optionsUse for L4 workloads, non-HTTP protocols
Internal TCP/UDPInternal passthrough Network Load BalancerPrivate load balancing inside VPC
GKE ingressGoogle Cloud load balancing via GKE Ingress/GatewayMatch with GKE architecture
Hybrid backend integrationLoad balancer with hybrid connectivity patterns where supportedConfirm backend type and connectivity requirement conceptually

Networking Traps

TrapCorrect thinking
“Cloud NAT lets users reach private VMs”Cloud NAT is outbound-only
“VPC peering creates a hub-and-spoke transit network”VPC peering is not transitive
“External IPs are required for Google API access”Private Google Access or Private Service Connect can avoid external IPs
“Firewall rules replace IAM”Network controls and identity controls solve different problems
“Global VPC means subnets are global”Subnets are regional
“Interconnect is encrypted by default like VPN”VPN provides encryption; for Interconnect, evaluate encryption requirements separately

Networking quick review

Networking questions often test whether you can keep traffic private, reduce latency, connect hybrid environments, or choose the right load balancer.

RequirementLikely Google Cloud concept
Isolate resourcesVPC, subnet design, firewall rules, IAM
Centralize networking for many projectsShared VPC
Private communication to Google APIsPrivate Google Access / Private Service Connect depending on use case
Private service publishing/consumptionPrivate Service Connect
Connect VPCsVPC Network Peering or Network Connectivity Center patterns
Outbound internet for private VMsCloud NAT
Hybrid connectivityCloud VPN or Cloud Interconnect
Low-latency dedicated hybrid linkCloud Interconnect
Global external HTTP(S) trafficGlobal external Application Load Balancer
Internal service trafficInternal load balancing
Edge cachingCloud CDN
DNS managementCloud DNS
DDoS/WAF protectionCloud Armor
Service-to-service control in GKEService mesh / network policies where appropriate

Networking traps

  • Cloud NAT is outbound only. It does not allow inbound internet connections to private instances.
  • Firewall rules are stateful, but you still need the right direction, target, source, and priority.
  • Shared VPC is commonly the right answer when central network teams manage networking for service projects.
  • VPC peering is not transitive. Do not assume A-to-B and B-to-C means A-to-C connectivity.
  • Private Google Access helps private VMs reach supported Google APIs without external IPs.
  • Cloud VPN is often faster to set up; Cloud Interconnect is typically used when dedicated, higher-throughput, lower-latency connectivity is required.
  • Use Cloud CDN and global load balancing for global web performance; do not solve every latency issue by adding more VMs.

IAM, Security, and Governance

Resource Hierarchy

LevelUse for
OrganizationEnterprise-wide policies, central governance
FolderDepartment, environment, business unit segmentation
ProjectWorkload/application boundary, billing/quota/IAM unit
ResourceService-specific access where appropriate
Notes and examples

IAM Reference

ConceptUsePCA distinction
PrincipalUser, group, service account, workload identityPrefer groups for humans and service accounts for workloads
RoleSet of permissionsUse predefined roles first; custom roles for least privilege gaps
Policy bindingGrants role to principal at resource scopeGrant at lowest practical scope
Basic rolesOwner, Editor, ViewerUsually too broad for production least privilege
Service accountApplication/workload identityDo not treat as a human user
Service account keyLong-lived credentialAvoid when federation or attached service accounts work
IAM ConditionsContext-aware conditional grantsUseful for time, resource, or attribute-based constraints
Deny policiesExplicitly restrict permissionsUse for guardrails where allow policies are too broad
Organization Policy ServiceConstraint-based governanceControls allowed configurations; not the same as IAM
Policy Controller / Config Controller patternsKubernetes/resource policy enforcementUseful in platform governance scenarios

Security Service Selection

RequirementPreferNotes
Centralized identity for workforceCloud Identity / external IdP federationUse SSO and groups for manageability
No service account keys from external workloadsWorkload Identity FederationReplaces long-lived keys for many external CI/CD and workload scenarios
GKE workloads use Google Cloud APIsWorkload Identity for GKEAvoid node-wide broad service account permissions
Secrets storageSecret ManagerDo not store secrets in images, source code, or metadata
Encryption key controlCloud KMS with CMEK where requiredDefault encryption exists, but CMEK gives customer-managed control
External key controlCloud External Key Manager where requiredUse only when scenario demands external key custody
Data perimeterVPC Service ControlsHelps reduce data exfiltration from supported services
App access without VPNIdentity-Aware ProxyUser/context-aware access to supported apps/resources
Threat/security postureSecurity Command CenterCentral visibility, findings, risk management
Asset inventoryCloud Asset InventoryAudit resources and IAM over time
Sensitive data inspectionSensitive Data ProtectionClassification, inspection, masking/tokenization use cases
Audit trailCloud Audit Logs, log sinksExport to BigQuery/Cloud Storage/SIEM as required

Security Traps

Scenario wordingWrong shortcutBetter answer
“Developers need temporary deployment access”Give OwnerUse least-privilege predefined/custom role, group, condition if applicable
“On-prem app needs Google Cloud access”Download service account keyPrefer Workload Identity Federation if supported
“Prevent data exfiltration from BigQuery/Cloud Storage”IAM onlyCombine IAM, VPC Service Controls, org policies, logging, DLP as needed
“KMS encrypts sensitive fields”Use KMS as classifierKMS manages keys; Sensitive Data Protection discovers/classifies/masks
“Network is private, so auth is unnecessary”Trust network boundary onlyUse identity, IAM, mTLS/service auth where appropriate
“Give GKE node service account broad access”Node-level permissionsUse Workload Identity for GKE per workload

Security and IAM quick review

Security scenarios often ask for the least privileged, most auditable, and most maintainable answer.

RequirementStrong approach
Human access controlIAM groups, predefined/custom roles, least privilege
Application identityService accounts
Avoid long-lived keysWorkload identity federation or attached service accounts instead of downloadable keys
Store secretsSecret Manager
Encryption key controlCloud KMS / Cloud HSM when appropriate
Central policy enforcementOrganization Policy Service
Detect security risksSecurity Command Center
Audit activityCloud Audit Logs
Restrict data exfiltration from supported servicesVPC Service Controls
Context-aware accessAccess Context Manager / BeyondCorp-style controls
Protect web appsCloud Armor
Private access to servicesPrivate Service Connect and private networking patterns

IAM decision rules

  1. Assign access to groups, not individual users, where possible.
  2. Use predefined roles before custom roles unless the scenario requires exact permission tailoring.
  3. Avoid Owner, Editor, and Viewer for production access unless the question explicitly justifies them.
  4. Use service accounts per workload rather than sharing one broad service account.
  5. Avoid service account keys when an identity-based alternative is available.
  6. Use organization policies to enforce constraints across projects.
  7. Use separation of duties for production deployments, security administration, and audit review.

Security traps

  • “Encrypt the data” is not enough; identify whether default encryption is sufficient or customer-managed keys are required.
  • Do not grant broad project-level roles to solve one resource-level permission issue.
  • Do not expose a private workload with public IPs when private connectivity is required.
  • Do not use firewall rules as the only security boundary for identity-sensitive services.
  • Do not confuse authentication with authorization.
  • Do not store API keys, passwords, or certificates in container images or source repositories.

Reliability, Availability, and Disaster Recovery

Core Terms

TermMeaningExam use
SLIMeasured indicator, such as latency or availabilityWhat you measure
SLOTarget for an SLIWhat you promise internally
SLAExternal contractual commitmentDo not invent values
Error budgetAllowed unreliability under SLOBalances release velocity and reliability
RTOMaximum acceptable recovery timeDrives failover automation
RPOMaximum acceptable data lossDrives replication/backup design
MTTRMean time to recoverImprove with automation and observability
Notes and examples

Availability Pattern Selection

RequirementPattern
Survive VM failureManaged instance group, health checks, autoscaling
Survive zone failureRegional managed instance group or regional managed service
Survive region failureMulti-region active-active or active-passive architecture
Low global latencyGlobal load balancing, CDN, regional backends near users
Stateless app HAMultiple zones/regions behind load balancer
Stateful relational HAManaged HA database option or replicated architecture
Backup/restore DRAutomated backups, tested restore procedures
Near-zero downtime deploymentsRolling, blue/green, or canary deployments
Fault isolationSeparate projects, regions, clusters, service boundaries

DR Decision Table

RTO/RPO pressureTypical architectureTradeoff
Low pressure, cost-sensitiveBackup and restoreCheapest, slowest recovery
ModeratePilot light / warm standbyFaster recovery, some idle cost
HighActive-passive with replicationMore automation and duplicate capacity
Very high global availabilityActive-active multi-regionHighest complexity; conflict/data design matters

Reliability Traps

TrapCorrect thinking
Backups are enough for all DRRestore time and data loss must meet RTO/RPO
Multi-zone equals multi-regionZonal failure protection is not regional disaster protection
Stateless autoscaling fixes database bottlenecksScale and protect the stateful tier separately
Active-active is always bestIt increases cost and complexity; use only when justified
Health checks alone guarantee app healthHealth checks must reflect real dependency readiness
Manual failover is acceptable for strict RTOAutomation is usually required for tight RTO

Application Modernization and Migration

StrategyMeaningGoogle Cloud fit
RehostMove as-isCompute Engine, Migrate to Virtual Machines, VMware Engine
ReplatformMinor changes to managed servicesCloud SQL, Cloud Run, GKE, managed load balancing
RefactorRedesign app architectureMicroservices, Pub/Sub, Cloud Run/GKE, managed databases
ReplaceAdopt SaaS or packaged solutionUse when custom app no longer provides business value
RetireRemove unused systemsCommon cost/risk reduction step
RetainKeep temporarilyUse for dependencies, compliance, or phased migration
Notes and examples

Migration Service Selection

NeedChoose
VM migrationMigrate to Virtual Machines
VMware environment migration with minimal changeGoogle Cloud VMware Engine
Database migrationDatabase Migration Service
Large online/offline object transferStorage Transfer Service
Physical appliance transfer for large datasetsTransfer Appliance
Streaming database change replicationDatastream
Landing zone setupResource hierarchy, IAM, Shared VPC, logging, billing, org policies
App dependency discoveryMigration assessment/discovery tooling and architecture review

Modernization Patterns

PatternUse whenReference design
Strangler figReplace legacy app graduallyRoute selected functions to new services while legacy remains
Event-driven decouplingSystems need async integrationPub/Sub between producers and consumers
API façadeLegacy backend needs modern interfaceAPI layer on Cloud Run/GKE with auth, logging, throttling
Containerize firstApp can run in containers but not fully refactoredCloud Run for stateless, GKE for Kubernetes/platform needs
Database offloadAnalytics overloads OLTP databaseReplicate/export to BigQuery

Migration and modernization review

Migration scenarios test sequencing, risk reduction, downtime tolerance, and data movement.

ScenarioStrong approach
Lift-and-shift VMsMigrate to Virtual Machines / Compute Engine migration tooling
Modernize monolith graduallyStrangler pattern, APIs, staged migration
Database migration with minimal downtimeDatabase Migration Service where supported, replication, cutover planning
Large object/data transferStorage Transfer Service, Transfer Appliance for very large/offline transfers
Hybrid period requiredCloud VPN or Cloud Interconnect, hybrid DNS, identity integration
Validate before cutoverPilot, test migration, dual-run, canary, rollback plan
Standardize infrastructureInfrastructure as code and repeatable deployment pipelines
Legacy app with minimal changesCompute Engine or VMware Engine depending on constraints

Migration traps

  • Replatforming everything at once increases risk unless the scenario requires it.
  • Choosing a solution that requires downtime when the requirement says minimal downtime.
  • Ignoring identity, DNS, firewall, monitoring, and rollback planning.
  • Moving data without validating consistency and application compatibility.
  • Forgetting that migration success includes operations after cutover, not only data transfer.

DevOps, CI/CD, and Infrastructure as Code

RequirementPreferNotes
Source repository integrationCloud Build with source triggers or existing CIMatch enterprise tooling if specified
Artifact storageArtifact RegistryContainer and package artifact management
Progressive delivery to GKE/Cloud RunCloud DeployManaged delivery pipelines
Infrastructure as codeTerraform, Infrastructure Manager, Config Controller where appropriatePCA tests concepts more than syntax
Policy as codeOrganization Policy, Policy Controller, CI checksEnforce guardrails before deployment
Secrets in CI/CDSecret Manager and workload identity/federationAvoid plaintext build variables
Deployment rollbackVersioned artifacts, blue/green/canary, release promotionDesign rollback before failure
Environment separationSeparate projects/folders per environmentEnables IAM, billing, quota, and blast-radius control
Notes and examples

Deployment Pattern Selection

PatternChoose whenTradeoff
Rolling updateNormal low-risk updatesSimple; some mixed-version period
Blue/greenFast rollback and environment validationDuplicate capacity during cutover
CanaryValidate with small traffic percentageRequires monitoring and traffic control
Immutable infrastructureAvoid configuration driftRequires automated build/deploy pipeline
Feature flagsDecouple release from deploymentRequires app-level implementation

Observability and Troubleshooting

NeedService / featureExam cue
Metrics and alertingCloud MonitoringSLOs, dashboards, uptime checks
LogsCloud LoggingCentral logs, log-based metrics
Audit trailCloud Audit LogsAdmin, data access, policy activity; verify what must be enabled/exported
Error aggregationError ReportingApp exceptions
Distributed tracingCloud TraceLatency across services
ProfilingCloud ProfilerCPU/memory performance over time
Log exportLog sinks to BigQuery, Cloud Storage, Pub/SubAnalytics, retention, SIEM integration
Resource inventoryCloud Asset Inventory“What changed?” and governance review
SLO managementCloud Monitoring SLOsReliability engineering scenarios
Notes and examples

Troubleshooting Decision Table

SymptomCheck firstLikely issue
403 permission deniedIAM binding scope, service account, org policy, VPC Service ControlsIdentity or policy problem
VM has no internet egressRoute, firewall egress, Cloud NAT, external IP, Private Google Access for APIsNetwork egress design
Private VM cannot call Google APIsPrivate Google Access / Private Service Connect, DNS, firewallMissing private API access path
GKE pod cannot access Google APIsWorkload Identity for GKE, Kubernetes service account binding, IAM roleWorkload identity mapping
High web latency globallyLoad balancer scope, CDN, backend region, database latencyRegional architecture for global users
Pub/Sub duplicatesSubscriber idempotency and ack handlingPub/Sub is at-least-once; design consumers accordingly
BigQuery query cost highPartitioning, clustering, selected columns, materialized views, reservations if applicableInefficient analytical query design
Cloud Run cold-start concernMin instances where justified, right runtime/container sizeServerless scaling tradeoff
Database overloaded by reportsRead replica or analytics offload to BigQueryMixing OLTP and analytics

Cost and Performance Optimization

AreaPractical controls
Compute EngineRightsizing, autoscaling, managed instance groups, committed use discounts where stable, Spot VMs for interruptible work
GKEAutopilot for reduced ops, right-sized requests/limits, cluster autoscaling, separate node pools for specialized workloads
Cloud RunConcurrency, min instances only when needed, efficient container startup, scale-to-zero for variable traffic
Cloud StorageLifecycle policies, storage class selection, avoid unnecessary duplication, monitor retrieval/egress patterns
BigQueryPartitioning, clustering, materialized views, avoid SELECT *, manage reservations/slots where appropriate
DatabasesRead replicas, caching, indexing, connection pooling, choose the right database model
NetworkingReduce unnecessary cross-region egress, use CDN for cacheable content, place compute near data/users
OperationsBilling export to BigQuery, budgets/alerts, labels, project separation, resource cleanup automation
Notes and examples

Cost Traps

TrapBetter approach
Use the cheapest VM shape for everythingMatch workload performance and reliability requirements
Use Spot VMs for critical stateful servicesUse Spot only when interruption is acceptable
Store all data in one class foreverApply lifecycle policies based on access
Run analytics on production OLTPOffload to BigQuery or replicas
Add min instances everywhereUse only where latency requirement justifies cost
Ignore network egressArchitect data locality and caching intentionally

Cost and performance review

PCA cost questions are usually about selecting an architecture that meets requirements without overengineering.

Cost areaReview points
ComputeAutoscaling, rightsizing, machine family selection, committed use discounts where appropriate
ServerlessPay-per-use can reduce idle cost, but understand concurrency and invocation patterns
StorageChoose appropriate storage class, lifecycle policies, retention, object versioning
NetworkingWatch egress, inter-region traffic, CDN caching, private connectivity needs
BigQueryPartitioning, clustering, materialized views, reservations/slots or on-demand fit
LoggingLog volume, exclusions, retention
DatabasesScale shape, replicas, HA, storage growth, backups
DRBalance RTO/RPO against standby capacity cost

Cost traps

  • Choosing active-active multi-region when the RTO/RPO does not require it.
  • Keeping development environments running 24/7 without need.
  • Ignoring network egress in distributed architectures.
  • Overprovisioning GKE clusters when serverless or autoscaling would satisfy the requirement.
  • Using premium managed services by default without matching workload needs.
  • Treating committed discounts as a first step before rightsizing and usage analysis.

Governance and Enterprise Landing Zone Checklist

DomainInclude in target architecture
Resource hierarchyOrganization, folders, projects by environment/app/team
IdentityCloud Identity or federation, groups, least privilege
NetworkingShared VPC, subnet strategy, hybrid connectivity, DNS, firewall policy
Security guardrailsOrg policies, IAM deny where appropriate, VPC Service Controls for sensitive services
LoggingCentral log sinks, audit log retention/export, SIEM integration if required
BillingBilling accounts, budgets, labels, export to BigQuery
Key managementKMS key rings/keys, CMEK requirements, separation of duties
CI/CDArtifact Registry, build/deploy pipelines, policy checks
MonitoringStandard dashboards, alerting, SLOs, incident response
Compliance evidenceAsset inventory, audit logs, security findings, documented controls

High-Yield Exam Distinctions

Do not confuseWithDistinction
Cloud SQLSpannerCloud SQL is regional managed relational; Spanner is distributed relational scale/global consistency
BigQueryCloud SQLBigQuery is analytics; Cloud SQL is transactional
BigtableBigQueryBigtable serves low-latency operational access; BigQuery analyzes large datasets
Pub/SubCloud TasksPub/Sub is event messaging; Cloud Tasks dispatches tasks to handlers with queue semantics
DataflowDataprocDataflow is managed Beam stream/batch; Dataproc is managed Spark/Hadoop
Cloud RunGKECloud Run is serverless containers; GKE is Kubernetes platform
GKE AutopilotGKE StandardAutopilot reduces node ops; Standard gives more control
Cloud NATLoad balancerNAT enables outbound private VM egress; load balancer handles inbound distribution
VPC peeringShared VPCPeering connects networks; Shared VPC centralizes network ownership for projects
IAMOrg PolicyIAM grants permissions; org policy constrains allowed configurations
KMSSecret ManagerKMS manages cryptographic keys; Secret Manager stores secrets
VPC Service ControlsFirewallVPC-SC protects service perimeters; firewall controls network traffic
Cloud CDNCloud StorageCDN caches/serves content closer to users; Storage holds objects

Scenario Answer Patterns

Scenario clueStrong candidate answer
“Reduce operational overhead”Prefer serverless/managed services; avoid self-managed clusters/VMs unless required
“Existing monolith on VMs, rapid migration”Rehost to Compute Engine or VMware Engine; modernize later
“Stateless container API with unpredictable traffic”Cloud Run behind appropriate load balancing/auth controls
“Kubernetes required by platform team”GKE Autopilot unless node-level control is required
“Global users, static and dynamic web content”Global external Application Load Balancer, Cloud CDN for cacheable assets, regional backends
“Strict least privilege”Groups, predefined/custom roles, minimal scope, service accounts, IAM Conditions if useful
“No public IPs on workloads”Private subnets, Cloud NAT for egress, Private Google Access/Private Service Connect
“Sensitive data perimeter”IAM plus VPC Service Controls, audit logs, org policies, DLP/Sensitive Data Protection
“Streaming sensor data”Pub/Sub ingestion, Dataflow processing, Bigtable or BigQuery depending on serving/analytics need
“Reports slow down production DB”Replicate/export to BigQuery; use read replicas when appropriate
“Need multi-region relational consistency”Cloud Spanner
“Need cache to reduce read latency”Memorystore in front of primary database
“Need central network administration”Shared VPC with host/service projects
“Need private hybrid connectivity”HA VPN for encrypted internet-based connectivity; Interconnect for private higher-throughput needs
“Need auditable infrastructure changes”IaC, CI/CD approvals, Cloud Audit Logs, Asset Inventory
Notes and examples

Common answer-choice patterns

If the option says…Be cautious because…
“Grant Owner role”Usually violates least privilege
“Create service account keys”Long-lived keys are often avoidable
“Use a single VM”May fail availability/scaling requirements
“Use BigQuery for app transactions”Wrong fit for OLTP
“Manually copy data”Weak for repeatability, scale, and auditability
“Expose the database publicly”Usually fails security requirements
“Use Cloud NAT for inbound access”Cloud NAT is outbound only
“Use multi-region everywhere”May overrun cost/complexity requirements
“Refactor all applications immediately”Often too risky for migration scenarios
“Rely only on backups for low RTO”Restore time may not meet requirement

Last-Minute PCA Review Checklist

  • Can you justify why a managed service is or is not suitable?
  • Can you separate OLTP, analytics, cache, object, block, and file storage?
  • Can you choose between Cloud Run, GKE, App Engine, and Compute Engine from requirements?
  • Can you explain Shared VPC, VPC peering, Private Service Connect, Private Google Access, Cloud NAT, VPN, and Interconnect?
  • Can you design for least privilege without using broad basic roles?
  • Can you identify when VPC Service Controls are needed and when they are not enough?
  • Can you map RTO/RPO to backup, warm standby, active-passive, or active-active?
  • Can you design centralized logging, monitoring, alerting, and audit export?
  • Can you optimize BigQuery, Cloud Storage, compute, and network egress without breaking requirements?
  • Can you spot answers that are technically possible but operationally heavy?

High-yield PCA mindset

The best answer usually balances business requirements, technical constraints, Google Cloud managed services, and operational simplicity.

Scenario clueArchitectural instinct
“Minimize operations”Prefer managed/serverless services over self-managed infrastructure
“Global users, low latency”Consider global load balancing, CDN, multi-region data, edge caching
“Strict RPO/RTO”Design for replication, automated failover, tested recovery, backups
“Regulated data”Use IAM least privilege, encryption, audit logging, VPC Service Controls where appropriate
“Existing on-premises integration”Consider Cloud VPN, Cloud Interconnect, hybrid DNS, identity federation
“Unpredictable traffic”Use autoscaling, serverless, queues, event-driven design
“Large analytics workloads”Think BigQuery, Dataflow, Dataproc, Pub/Sub, Cloud Storage
“Transactional relational workload”Consider Cloud SQL, AlloyDB, or Spanner depending on scale and global needs
“Need portability / Kubernetes standardization”Consider Google Kubernetes Engine, but weigh operational complexity
“Cost is a key constraint”Rightsize, autoscale, choose storage classes, avoid unnecessary egress, use commitments carefully

Fast architecture decision flow

    flowchart TD
	    A[Read business goal] --> B[Identify constraints]
	    B --> C{Primary workload type?}
	    C -->|Web/API/app| D[Choose compute pattern]
	    C -->|Data/analytics| E[Choose storage and pipeline]
	    C -->|Hybrid/migration| F[Choose connectivity and migration path]
	    D --> G[Secure identity, network, secrets]
	    E --> G
	    F --> G
	    G --> H[Design reliability and DR]
	    H --> I[Optimize cost and operations]
	    I --> J[Select answer with least complexity that satisfies requirements]

Core Google Cloud resource model

Understand the hierarchy because many PCA security and governance questions depend on it.

LevelWhat to remember
OrganizationTop-level boundary for centralized policy, IAM, billing, and governance
FolderGroups projects by department, environment, app, or business unit
ProjectMain isolation boundary for resources, IAM, quotas, billing attribution, APIs
ResourceIndividual service objects such as VM instances, buckets, clusters, datasets

Common hierarchy traps

  • Granting permissions too broadly at the organization level when project-level or resource-level access is enough.
  • Using primitive roles when predefined or custom roles better support least privilege.
  • Forgetting that inherited IAM can create access unintentionally.
  • Mixing production and non-production workloads in the same project without a strong reason.
  • Ignoring organization policies when the scenario asks for centralized enforcement.

Compute and application platform review

Choose compute by matching operational burden, portability, scaling, and runtime requirements.

Service / patternBest fitWatch for
Cloud RunStateless containers, APIs, event-driven apps, scale-to-zero patternsNot ideal for highly stateful workloads without external state storage
Cloud FunctionsLightweight event-driven functionsAvoid when container control or complex runtime dependencies are required
App EngineManaged web apps with minimal infrastructure managementLess control than GKE or Compute Engine
Google Kubernetes EngineContainer orchestration, microservices, Kubernetes portability, platform teamsMore operational complexity than Cloud Run
Compute EngineCustom VMs, legacy workloads, special OS/configuration needsRequires more management than serverless/managed platforms
Managed instance groupsAutoscaled VM fleets, rolling updates, self-healingDesign health checks and load balancing correctly
BatchBatch jobs and high-throughput non-interactive workloadsNot for low-latency request/response apps
Cloud Run jobsContainerized jobs without always-on infrastructureGood for task execution, not long-lived services
VMware EngineVMware-based migration or extension scenariosUse when VMware compatibility matters, not as default modernization
Notes and examples

Compute decision rules

  • Prefer Cloud Run for stateless containerized services when you want managed scaling and low operational overhead.
  • Prefer GKE when Kubernetes control, service mesh, custom networking, or workload portability is a requirement.
  • Prefer Compute Engine when the workload needs custom machine configuration, legacy software, or VM-level control.
  • Prefer managed instance groups over unmanaged VM fleets for autoscaling and self-healing.
  • Externalize state to managed storage such as Cloud SQL, Spanner, Firestore, Memorystore, Cloud Storage, or Bigtable.

Container and delivery topics to review

NeedGoogle Cloud option
Store container imagesArtifact Registry
Build/test/deploy pipelineCloud Build, Cloud Deploy
Kubernetes workload identityWorkload Identity / workload identity federation patterns
Enforce trusted container deploymentBinary Authorization
Manage GKE configurationConfig Sync / policy-based controls where appropriate
Canary or progressive deliveryCloud Deploy, load balancing, GKE rollout strategies, Cloud Run revisions
SecretsSecret Manager, not hardcoded environment values

BigQuery and analytics review

BigQuery questions often mix cost, performance, governance, and ingestion choices.

NeedReview point
Analyze large datasets with SQLBigQuery
Load data periodicallyBatch loads, BigQuery Data Transfer Service, Storage Transfer Service depending on source
Stream eventsPub/Sub into Dataflow or supported streaming ingestion patterns
Reduce query costPartitioning, clustering, selective columns, materialized views, query optimization
Control data accessDataset/table permissions, authorized views, row-level and column-level controls where appropriate
Separate storage and computeUnderstand BigQuery’s serverless analytics model
Share governed datasetsUse appropriate IAM and data sharing patterns

BigQuery traps

  • Querying unnecessary columns increases scanned data when using on-demand pricing models.
  • Partitioning is useful only when queries filter on partitioned fields.
  • Clustering helps when queries filter or aggregate on clustered columns.
  • BigQuery is not a replacement for a low-latency transactional application database.
  • Governance is not just IAM; consider dataset structure, views, policy tags, and auditability.

Reliability, DR, and SRE review

PCA reliability questions usually test whether the architecture meets availability and recovery requirements without unnecessary complexity.

Key terms:

  • RTO: how quickly the service must be restored.
  • RPO: how much data loss is acceptable.
  • SLO: target reliability level for a service.
  • Error budget: allowable unreliability over a period.
Notes and examples\[ \text{Availability} = \frac{\text{Successful requests}}{\text{Total valid requests}} \]\[ \text{Error budget} = 1 - \text{SLO} \]

Reliability design table

RequirementDesign instinct
Survive zonal failureUse regional managed services or deploy across zones
Survive regional failureUse multi-region or cross-region DR design
Low RTOAutomate failover and keep warm/active capacity
Low RPOUse synchronous or near-real-time replication where appropriate
Stateless web tierUse load balancing and autoscaling across zones
Stateful databaseUse managed HA, replicas, backups, and tested restore
Message bufferingPub/Sub or queues to absorb spikes and decouple systems
Graceful degradationCache, fallback responses, circuit breakers
Operational visibilityCloud Monitoring, Cloud Logging, tracing, alerts, dashboards

DR pattern comparison

PatternUse whenTradeoff
Backup and restoreLower cost, higher RTO acceptableSlowest recovery
Pilot lightMinimal standby environmentFaster than restore, still needs scale-up
Warm standbyReduced-capacity duplicate environmentHigher cost, faster recovery
Active-activeVery low RTO, global availabilityHighest complexity and cost

Reliability traps

  • Backups do not prove recoverability; restores must be tested.
  • Multi-zone is not the same as multi-region.
  • Autoscaling does not fix database bottlenecks automatically.
  • A load balancer cannot route to healthy capacity if the backend design is single-zone.
  • SLOs should reflect user experience, not only infrastructure uptime.
  • Alerting on every metric creates noise; alert on symptoms and user-impacting conditions.

Operations and observability

Operational questions look for maintainable, automated, auditable solutions.

NeedGoogle Cloud capability
Metrics and alertingCloud Monitoring
LogsCloud Logging
Distributed tracesCloud Trace
Error aggregationError Reporting
Service health and SLOsCloud Monitoring service monitoring
Audit trailCloud Audit Logs
Deployment automationCloud Build, Cloud Deploy, Terraform/infrastructure as code patterns
Config governanceOrganization policies, policy-as-code, Config Controller/Config Sync where appropriate
Incident responseDashboards, alerts, runbooks, postmortems

Operations traps

  • Manual console changes are rarely the best answer for repeatable production architecture.
  • Alerts should be actionable and tied to service impact.
  • Logs have cost and retention implications; collect what you need and manage retention.
  • Monitoring must include dependencies such as databases, queues, and external services.
  • Production changes should be versioned, reviewed, and reversible.

Scenario-reading technique for PCA

Use this checklist on every long question.

  1. Identify the primary business goal. Cost reduction, availability, compliance, modernization, speed, scalability, or migration?
  2. Underline hard constraints. “Must,” “without downtime,” “least operational overhead,” “existing Kubernetes,” “global,” “private,” “auditable.”
  3. Classify the workload. Web app, batch, analytics, streaming, relational transaction, document, ML, hybrid.
  4. Choose managed services first. Only add lower-level infrastructure when needed.
  5. Apply security by default. IAM least privilege, private access, encryption, audit logs, secrets management.
  6. Check reliability. Does the answer meet stated RTO/RPO and failure-domain requirements?
  7. Check cost and simplicity. Avoid architectures that exceed the requirement.
  8. Eliminate answers with one fatal flaw. Broad IAM, public exposure, manual process, single point of failure, wrong data model.

Final rapid review checklist

Before starting original practice questions, confirm that you can quickly explain:

  • When to choose Cloud Run, GKE, Compute Engine, App Engine, or Cloud Functions
  • How Shared VPC, VPC peering, Private Google Access, Private Service Connect, Cloud NAT, Cloud VPN, and Cloud Interconnect differ
  • When to choose Cloud SQL, AlloyDB, Spanner, Firestore, Bigtable, BigQuery, or Cloud Storage
  • How IAM inheritance, service accounts, least privilege, and organization policies work
  • How to design for zonal, regional, and multi-region resilience
  • How RTO, RPO, SLOs, and error budgets influence architecture choices
  • How to reduce BigQuery, compute, storage, logging, and network costs
  • How to plan staged migrations with validation and rollback
  • How to spot overbroad permissions, public exposure, manual operations, and single points of failure

Connect this review to practice

After this Cheat Sheet, move into IT Mastery practice:

  • Use topic drills for weak areas such as IAM, networking, data services, and reliability.
  • Use original practice questions to build scenario-reading speed.
  • Use a full question bank to compare similar services under realistic constraints.
  • Review detailed explanations carefully, especially when an answer is technically possible but not the best architectural choice.

Practical next step: choose one weak domain, complete a focused topic drill, and review every explanation until you can state why the correct answer is better than the plausible distractors.

Put the review into practice