GCP Professional Cloud Architect — Ultimate Short Notes (2026)
Single source of truth. Covers all 6 exam domains, 4 case studies, Well-Architected Framework, decision trees, tips, tricks, and edge cases. Each section ends with real-style practice questions.
TABLE OF CONTENTS
- Exam Overview & Strategy
- Domain 1 — Design & Plan a Cloud Solution Architecture
- Domain 2 — Manage & Provision Cloud Solution Infrastructure
- Domain 3 — Design for Security & Compliance
- Domain 4 — Analyze & Optimize Technical & Business Processes
- Domain 5 — Manage Implementations of Cloud Architecture
- Domain 6 — Ensure Solution & Operations Excellence
- Well-Architected Framework
- Case Studies Deep Dive
- Decision Trees & Cheat Sheets
- Critical Edge Cases & Exam Traps
1. EXAM OVERVIEW & STRATEGY
| Item | Detail |
|---|---|
| Questions | 50-60 multiple choice + multi-select |
| Duration | 2 hours |
| Case studies | 2 per exam (from pool of 4), shown on split screen |
| Case study weight | 20-30% of questions |
| Passing score | ~80% (not officially published) |
| Cost | $200 |
Strategy Tips:
- Case studies appear on a split screen — skim them first (2-3 min each), then answer related questions
- Multi-select questions say “Select TWO” — always select exactly the count asked
- When two answers are both correct, pick the one that is most managed, least operational overhead, most Google-native
- “Minimize cost” ≠ cheapest option — it means cheapest that still meets ALL requirements
- “Minimize operational overhead” = fully managed service wins (e.g., Cloud SQL > self-managed MySQL on GCE)
- Read ALL four options before answering — trap answers are plausible
- Flag long questions, answer easy ones first, return later
2. DOMAIN 1 — DESIGN & PLAN A CLOUD SOLUTION ARCHITECTURE
2.1 Compute — Which Service When?
| Service | Use When | Key Limits / Notes |
|---|---|---|
| Compute Engine (GCE) | Full OS control, GPU/TPU, lift-and-shift, legacy apps, licensing | Sole-tenant nodes for licensing; preemptible/spot VMs for batch (24h max, can be preempted) |
| GKE (Google Kubernetes Engine) | Containerized microservices, portability, hybrid (Anthos) | Autopilot = fully managed nodes; Standard = you manage node pools |
| Cloud Run | Stateless containers, event-driven, scale-to-zero, HTTP services | Max 60min request timeout; max 32 GiB memory; concurrency up to 1000 |
| Cloud Functions | Single-purpose event triggers, lightweight glue code | Gen2 built on Cloud Run; 60min timeout (gen2); 32 GiB memory (gen2) |
| App Engine Standard | Simple web apps, auto-scale to zero, fast deploys | Limited runtimes, no custom binaries, no background processes (use Standard for scale-to-zero) |
| App Engine Flexible | Custom runtimes, Docker containers, WebSockets, SSH | Does NOT scale to zero (min 1 instance), costs more than Standard |
| Cloud Batch | HPC, batch jobs with resource allocation | GPU support, job queuing, good for genomics/rendering |
| Bare Metal Solution | Oracle DB, SAP HANA, specialized hardware | Runs in Google’s data center extension, connected via Partner Interconnect |
Tips & Edge Cases — Compute
- “Migrate VM as-is” → Compute Engine (lift-and-shift with Migrate to VMs)
- “Containerize VM” → Migrate to Containers, then GKE
- “Scale to zero + container” → Cloud Run (NOT App Engine Flex)
- Preemptible vs Spot VMs: Spot is the newer name; same concept (can be reclaimed); use for fault-tolerant batch workloads
- Sole-tenant nodes: Required for BYOL (bring your own license) and workload isolation compliance
- Custom machine types: Can save 40%+ vs standard sizes; use when workload needs unusual CPU:memory ratio
- Committed Use Discounts (CUDs): 1 or 3 year; resource-based CUDs (vCPU/memory) or spend-based CUDs ($$)
- Sustained Use Discounts (SUDs): Automatic, no commitment; up to 30% for instances running 25%+ of month
- GKE Autopilot: Google manages nodes, you only define pods → less operational overhead, pod-level billing
- GKE Standard: You manage node pools, can use preemptible nodes, more control
2.2 Storage — Which Service When?
| Service | Type | Use When | Key Notes |
|---|---|---|---|
| Cloud Storage (GCS) | Object | Unstructured data, backups, static hosting, data lake | 4 classes: Standard, Nearline (30d), Coldline (90d), Archive (365d); lifecycle policies auto-transition |
| Persistent Disk | Block | VM boot/data disks | pd-standard (HDD), pd-balanced (SSD), pd-ssd (SSD), pd-extreme (high IOPS); regional PD for HA |
| Filestore | File (NFS) | Shared file system across VMs/GKE | Basic, High Scale, Enterprise tiers; Enterprise for cross-zone HA |
| Local SSD | Block (ephemeral) | Ultra-low latency scratch space | Data lost on VM stop/preemption; 375 GB each, up to 24 per VM |
Cloud Storage Classes — Decision Matrix
| Class | Min Storage Duration | Access Pattern | Use Case |
|---|---|---|---|
| Standard | None | Frequent | Hot data, serving, analytics input |
| Nearline | 30 days | ~Once/month | Backups, long-tail content |
| Coldline | 90 days | ~Once/quarter | DR copies, archival |
| Archive | 365 days | ~Once/year | Compliance archives, rarely accessed |
Trick: Autoclass automatically moves objects between classes based on access patterns — use when access patterns are unpredictable.
Edge Cases:
- Object Lifecycle Management: Can delete, change class, but cannot move TO a warmer class via lifecycle; only colder
- Bucket Lock for retention: cannot be removed once set (compliance). Retention policy without lock can be removed
- Signed URLs: Temporary access without Google account; time-limited
- Requester Pays: Useful when sharing large datasets; requestor pays egress
- Dual-region / multi-region: Higher availability + data residency + faster cross-region access
2.3 Databases — The Critical Decision Tree
Need relational?
├── YES
│ ├── Need global scale / 99.999% SLA? → Cloud Spanner
│ ├── Need PostgreSQL compatibility + high perf? → AlloyDB
│ ├── Standard RDBMS (MySQL/PostgreSQL/SQL Server)? → Cloud SQL
│ └── Need full control / custom config? → Self-managed on GCE
├── NO → What data model?
│ ├── Key-value, massive scale, time-series, IoT → Cloud Bigtable
│ ├── Document (mobile/web, real-time sync) → Firestore
│ ├── In-memory cache/sessions → Memorystore (Redis/Memcached)
│ ├── Graph → JanusGraph on Bigtable or use Spanner Graph
│ └── Analytics warehouse → BigQuery
| Database | Key Differentiator | SLA | Edge Case to Remember |
|---|---|---|---|
| Cloud SQL | Managed MySQL/PostgreSQL/SQL Server | 99.95% (HA) | Max 64 TB storage; read replicas cross-region; automated backups |
| Cloud Spanner | Global relational, horizontal scaling | 99.999% (multi-region) | Minimum cost ~$0.90/hr per node; use for financial, inventory, gaming leaderboards |
| AlloyDB | PostgreSQL-compatible, columnar engine | 99.99% | 4x faster than standard PostgreSQL for transactional, 100x for analytical queries |
| Cloud Bigtable | Wide-column NoSQL, petabyte-scale | 99.99% (replicated) | Single-row key design is critical; NOT for small datasets (<1TB); good for IoT, time-series |
| Firestore | Document NoSQL, real-time sync | 99.999% (multi-region) | Native mode (new apps) vs Datastore mode (legacy); cannot switch modes after creation |
| Memorystore | Redis/Memcached managed cache | 99.9% | Use for session caching, leaderboards, pub/sub light; Redis supports persistence |
| BigQuery | Serverless analytics warehouse | 99.99% | On-demand vs flat-rate (editions) pricing; streaming inserts vs batch loads |
| Firebase Realtime DB | JSON tree, mobile real-time sync | - | Simpler than Firestore; use for small mobile apps only |
Tips:
- “ACID + global” → Spanner
- “PostgreSQL compatible + analytical queries” → AlloyDB
- “IoT sensor data at scale” → Bigtable
- “Real-time sync to mobile” → Firestore (native mode)
- “Sub-millisecond cache” → Memorystore
- “Analytics on petabytes” → BigQuery
- Cloud SQL HA: Multi-zone with automatic failover; regional persistent disk
- Spanner vs Cloud SQL: Spanner for global scale and 99.999%; Cloud SQL for lower cost standard workloads
2.4 Networking
Core Networking Concepts
| Concept | Key Points |
|---|---|
| VPC | Global resource; subnets are regional; auto-mode (1 subnet/region) vs custom-mode (you define) |
| Shared VPC | Central network management; host project owns VPC, service projects use it; good for org-wide networking |
| VPC Peering | Connects 2 VPCs directly; non-transitive (A↔B, B↔C does NOT give A↔C); no overlapping CIDRs |
| Firewall Rules | Priority 0 (highest) to 65535 (lowest); default deny-ingress, allow-egress; use tags/service accounts |
| Cloud NAT | Managed NAT for private VMs to reach internet; no need for public IPs |
| Private Google Access | VMs with internal IPs only can reach Google APIs; enabled per subnet |
| Private Service Connect | Private endpoints for Google services + third-party; uses internal IPs |
| Serverless VPC Access | Cloud Run / Cloud Functions → VPC resources (Memorystore, Cloud SQL private IP) |
Hybrid/Multi-cloud Connectivity
| Method | Bandwidth | Latency | Use When |
|---|---|---|---|
| Cloud VPN (HA VPN) | Up to 3 Gbps per tunnel | Internet-based | Dev/test, small workloads, encrypted over internet |
| Dedicated Interconnect | 10 or 100 Gbps per link | Low, private | High bandwidth, colocation facility is available |
| Partner Interconnect | 50 Mbps–50 Gbps | Low-medium | No colocation; go through service provider |
| Cross-Cloud Interconnect | 10 or 100 Gbps | Low | GCP ↔ AWS/Azure/Oracle direct |
| Cloud Router | - | - | Dynamic routing (BGP) for all interconnect/VPN options |
Shared VPC vs VPC Peering — Decision
| Criteria | Shared VPC | VPC Peering |
|---|---|---|
| Ownership | Single org, centralized network admin | Any 2 VPCs, even cross-org |
| Transitivity | Transitive via host project | Non-transitive |
| Subnet sharing | Yes | No (separate subnets) |
| Firewall management | Centralized | Per-VPC |
| Best for | Enterprise with central network team | Partner integrations, separate teams |
Load Balancers — Which One?
| Type | Global/Regional | Layer | Protocol | Use When |
|---|---|---|---|---|
| External Application LB | Global | 7 | HTTP(S) | Web apps, URL-based routing, CDN, Cloud Armor |
| Internal Application LB | Regional | 7 | HTTP(S) | Internal microservices |
| External Network LB (proxy) | Global | 4 | TCP/SSL | TCP proxy, SSL offload |
| External Network LB (passthrough) | Regional | 4 | TCP/UDP | UDP, non-HTTP, IP preservation |
| Internal Network LB (passthrough) | Regional | 4 | TCP/UDP | Internal TCP/UDP backends |
| Cross-region Internal LB | Global | 4/7 | TCP/HTTP(S) | Multi-region internal traffic |
Key Tip: If the question says “global HTTP(S) load balancing” → External Application LB. If “internal services” → Internal Application LB or Internal Network LB.
DNS
- Cloud DNS: Managed authoritative DNS; public and private zones
- Private zones: Resolve internal names within VPC; can share across VPCs via DNS peering
- DNS peering: Forward queries from one VPC to another’s private zone
- Hybrid DNS: Use Cloud DNS inbound/outbound forwarding for on-prem ↔ GCP resolution
CDN & Edge
- Cloud CDN: Cache content at Google’s edge; works with External Application LB; supports cache invalidation
- Media CDN: For large-scale media delivery (video streaming)
- Cloud Armor: WAF + DDoS protection; attach to External Application LB; rate limiting, geo-blocking, preconfigured WAF rules (OWASP top 10)
2.5 Data & Analytics
| Service | Purpose | Key Notes |
|---|---|---|
| BigQuery | Serverless data warehouse | SQL-based; on-demand or editions (slot-based) pricing; supports ML (BQML), BI Engine (fast cache), materialized views |
| Dataflow | Stream + batch ETL (Apache Beam) | Fully managed; auto-scaling; exactly-once processing; templates for common patterns |
| Dataproc | Managed Spark/Hadoop | For existing Spark/Hadoop workloads; ephemeral clusters to save cost; use for migration from on-prem Hadoop |
| Cloud Composer | Managed Apache Airflow | Workflow orchestration; DAGs for complex pipelines; NOT for data processing itself |
| Pub/Sub | Messaging/event streaming | At-least-once delivery; push & pull; ordering with ordering keys; exactly-once via Dataflow |
| Pub/Sub Lite | Budget messaging | Zonal (not global); lower cost, you manage capacity; for high-volume, cost-sensitive streaming |
| Datastream | Change data capture (CDC) | Oracle/MySQL → BigQuery/Cloud Storage replication; serverless |
| Data Fusion | Visual ETL (CDAP-based) | Code-free data integration; for citizens/analysts; more expensive than Dataflow |
| Dataplex | Data mesh / governance | Discover, organize, govern data across lakes/warehouses/marts |
| Looker / Looker Studio | BI & visualization | Looker: enterprise BI with LookML; Looker Studio: free dashboards |
| BigQuery Omni | Multi-cloud analytics | Query data in AWS S3 / Azure Blob without moving it |
Tips:
- “Existing Spark jobs” → Dataproc
- “New ETL pipeline, no existing code” → Dataflow
- “Orchestrate multiple jobs” → Cloud Composer
- “Real-time streaming ingest” → Pub/Sub → Dataflow → BigQuery
- “CDC from Oracle” → Datastream
- “Query data in S3 without moving” → BigQuery Omni
- BigQuery streaming inserts cost more but are real-time; batch loads from GCS are free
- BigQuery partitioning (time, integer, ingestion time) + clustering (up to 4 columns) = performance + cost savings
2.6 AI/ML & Generative AI
| Service | Use Case | Key Note |
|---|---|---|
| Vertex AI | End-to-end ML platform | Training, tuning, serving, pipelines, experiments, feature store |
| AutoML (via Vertex AI) | No-code model training | Image, text, tabular, video classification; minimal ML expertise needed |
| BigQuery ML | SQL-based ML | Train models inside BigQuery without data movement; good for analysts |
| Pre-trained APIs | No training needed | Vision, Speech-to-Text, Natural Language, Translation, Video Intelligence |
| Vertex AI Gemini | Generative AI | Enterprise GenAI with data governance; VPC SC compatible |
| Model Garden | Pre-trained model catalog | Deploy open-source & Google models on Vertex AI |
| Vertex AI Search | Enterprise search | RAG (retrieval-augmented generation) over your data |
| Vertex AI Agents | AI agents | Build conversational agents grounded in your data |
| Document AI | Document processing | OCR + extraction for invoices, receipts, forms |
Decision Flow:
Need AI/ML?
├── Have labeled data + ML expertise? → Vertex AI custom training
├── Have labeled data + NO ML expertise? → AutoML
├── Need SQL-based ML, data already in BigQuery? → BigQuery ML
├── Pre-built capability (OCR, sentiment, STT)? → Pre-trained APIs
├── Need GenAI (summarize, generate, chat)? → Vertex AI Gemini
└── Need enterprise search over documents? → Vertex AI Search
Domain 1 — Practice Questions
Q1. A global e-commerce company needs a relational database for their product catalog that must handle 100,000+ reads/sec globally, support transactions, and maintain 99.999% availability. The team currently uses MySQL. Which service is most appropriate?
A. Cloud SQL with read replicas in every region B. Cloud Spanner with a multi-region configuration C. AlloyDB with cross-region replication D. Cloud Bigtable with a composite row key
Answer: B Why B: Spanner provides global horizontal scaling, strong consistency, and a 99.999% SLA (multi-region). Why not A: Cloud SQL read replicas are eventually consistent and don’t provide 99.999% SLA globally. Why not C: AlloyDB is regional (99.99%). Why not D: Bigtable is NoSQL with no ACID transactions.
Q2. A media company has 500 TB of video files accessed frequently for the first week, then rarely for 1 year, and must retain 7 years for compliance. How should they store this data most cost-effectively?
A. Store in Standard class with lifecycle policy to Nearline after 7 days, Coldline after 90 days, Archive after 365 days B. Store everything in Archive class from day one C. Store on Persistent Disks attached to Compute Engine D. Store in Filestore Enterprise
Answer: A Why A: Lifecycle policies automate class transitions matching the access pattern. Why not B: Archive has a 365-day minimum and high retrieval cost—week-1 access would be extremely expensive. Why not C/D: Block/file storage costs far more than object storage.
Q3. Your company requires all traffic between on-premises and GCP to be encrypted and private, with consistent bandwidth of 50 Gbps for a data migration project. You have space in a colocation facility near a Google peering edge. What should you use?
A. HA VPN with multiple tunnels B. Dedicated Interconnect with IPsec encryption C. Partner Interconnect D. Transfer Appliance
Answer: B Why B: Dedicated Interconnect provides up to 100 Gbps with low latency via colocation; adding MACsec or VPN over it provides encryption. Why not A: VPN max ~3 Gbps per tunnel—even 17 tunnels won’t reliably deliver 50 Gbps. Why not C: Partner Interconnect is for when you DON’T have colocation. Why not D: Transfer Appliance is for offline bulk data transfer, not continuous connectivity.
Q4. A startup wants to deploy a containerized REST API that receives a few hundred requests per day during business hours and zero traffic at night. They want to minimize cost. Which compute option is best?
A. GKE Standard with preemptible nodes B. Cloud Run C. App Engine Flexible D. Compute Engine with autoscaler
Answer: B Why B: Cloud Run scales to zero → no cost during idle time; ideal for bursty, low-traffic HTTP workloads. Why not A: GKE has minimum node costs even idle. Why not C: App Engine Flex doesn’t scale to zero (min 1 instance). Why not D: CE autoscaler has min 1 VM.
Q5. A team needs to migrate their on-premises Hadoop cluster (running Spark jobs nightly, processing 10 TB). They want to minimize refactoring and reduce costs. What approach is recommended? (Select TWO)
A. Use Dataproc with ephemeral clusters that spin up for each job B. Rewrite all Spark jobs as Dataflow pipelines C. Store data in Cloud Storage instead of HDFS D. Run a permanent Dataproc cluster 24/7
Answer: A, C Why A: Ephemeral clusters = pay only when running jobs. Why C: GCS as the storage layer decouples storage from compute, enables cluster shutdown between jobs. Why not B: “Minimize refactoring” — rewriting all jobs is maximal refactoring. Why not D: 24/7 cluster eliminates cost savings of cloud migration.
3. DOMAIN 2 — MANAGE & PROVISION CLOUD SOLUTION INFRASTRUCTURE
3.1 Resource Hierarchy
Organization (org)
└── Folders (optional, nestable)
└── Projects
└── Resources
- Organization: Top-level; tied to Cloud Identity / Google Workspace domain
- Folders: Group projects by dept / env / team; inherit policies from parent
- Projects: Billing boundary; all resources live in a project
- Policy inheritance: Policies set at org/folder level cascade DOWN; child CANNOT remove parent restriction (deny always wins)
Organization Policies (Key Ones)
| Policy | Purpose |
|---|---|
constraints/compute.vmExternalIpAccess |
Deny external IPs on VMs |
constraints/iam.allowedPolicyMemberDomains |
Restrict IAM to specific domains only |
constraints/compute.restrictSharedVpcSubnetworks |
Limit which subnets service projects can use |
constraints/gcp.resourceLocations |
Restrict where resources can be created (regions) |
constraints/compute.requireShieldedVm |
Enforce Shielded VMs only |
constraints/sql.restrictPublicIp |
Prevent Cloud SQL public IPs |
Tip: Org policies are the first line of governance. If a question says “prevent teams from creating public-IP VMs org-wide,” the answer is org policy constraint, not IAM.
3.2 Infrastructure as Code (IaC)
| Tool | Use When |
|---|---|
| Terraform | Multi-cloud IaC; declarative; state management; modules; most versatile |
| Deployment Manager | GCP-native IaC; YAML/Jinja/Python templates; DEPRECATED — prefer Terraform |
| Config Connector | Manage GCP resources via Kubernetes CRDs; use when GKE is your control plane |
| Pulumi | IaC with general-purpose languages (Python, Go, etc.) |
Edge Case: If the question mentions “Deployment Manager” as an option and “Terraform” is also there, Terraform is preferred unless the question specifically asks for a Google-native tool or existing DM templates.
3.3 Billing & Cost Management
| Concept | Key Points |
|---|---|
| Billing Account | Linked to projects; organization can have multiple; one billing account can fund many projects |
| Budgets & Alerts | Set $ thresholds; alerts via email, Pub/Sub (auto-action), or programmatic |
| Billing Export | Export to BigQuery for analysis; standard, detailed, or pricing export |
| Labels | Key-value metadata on resources; use for cost allocation by team/env/app |
| Committed Use Discounts | 1yr or 3yr; resource-based (vCPU/RAM) or spend-based ($$); up to 57% savings |
| Sustained Use Discounts | Automatic, no commitment; up to 30%; applies to GCE, GKE Standard nodes |
| Spot/Preemptible VMs | Up to 91% discount; can be reclaimed any time; for fault-tolerant batch |
| Recommender | Automated right-sizing recommendations; idle resource identification |
| Active Assist | Suite of recommendation tools (firewall insights, IAM recommender, etc.) |
Tip: “Reduce cost without affecting production” → Committed Use Discounts for steady-state + Spot VMs for batch. “Identify waste” → Recommender + billing export to BigQuery.
3.4 Landing Zone Design
A landing zone is the foundational cloud environment for an enterprise:
- Resource hierarchy (org → folders → projects)
- Identity (Cloud Identity, GCDS for AD sync, federation via SAML)
- Networking (Shared VPC, hub-and-spoke or similar)
- Security (org policies, security guardrails)
- Logging (centralized audit logs)
- Billing (billing account structure, budgets)
Cloud Foundation Toolkit: Terraform modules for deploying opinionated landing zones.
Domain 2 — Practice Questions
Q6. A large enterprise wants to ensure no team can create Compute Engine instances with external IP addresses, across all 50+ projects. Some projects don’t exist yet. What is the most effective approach?
A. Create a custom IAM role that lacks compute.instances.create permission B. Set an Organization Policy constraint on the organization node to deny VM external IPs C. Write a Cloud Function that monitors instance creation and deletes non-compliant VMs D. Document a policy and require teams to review it quarterly
Answer: B
Why B: Organization policy compute.vmExternalIpAccess at org level is inherited by ALL current and future projects; it’s preventive, not reactive. Why not A: Removing create permission entirely prevents creating VMs at all, not just restricting IPs. Why not C: Reactive — VMs briefly exist with public IPs. Why not D: Not enforceable technically.
Q7. Your company has dev, staging, and production environments. You need to centralize network management while allowing each team to manage their own resources. How should you structure this?
A. One project per environment, each with its own VPC B. Shared VPC with a host project for networking; separate service projects for dev, staging, prod C. VPC Peering between all project VPCs D. One project for everything with IAM roles to separate access
Answer: B Why B: Shared VPC centralizes network in the host project while allowing teams autonomy in service projects. Why not A: Decentralized networking = inconsistent policies. Why not C: VPC peering is non-transitive and doesn’t centralize management. Why not D: One project is an anti-pattern for isolation.
Q8. You need to predict your cloud spend for the next 3 years, track costs per department, and automate alerts when spending exceeds budgets. What combination should you use? (Select TWO)
A. Apply labels to all resources by department and export billing to BigQuery B. Use Cloud Monitoring custom dashboards for cost tracking C. Set up budgets with Pub/Sub notifications to trigger automated responses D. Use Cloud Logging to track API calls related to spending
Answer: A, C Why A: Labels + BigQuery export enables granular cost analysis by department. Why C: Budgets with Pub/Sub enable automated alerts and actions (e.g., disable billing, notify Slack). Why not B: Cloud Monitoring dashboards don’t natively handle billing data. Why not D: Logging tracks API activity, not spend.
4. DOMAIN 3 — DESIGN FOR SECURITY & COMPLIANCE
4.1 Identity & Access Management (IAM)
IAM Model
WHO (Member/Principal) + WHAT (Role) + WHERE (Resource) = IAM Policy Binding
| Concept | Details |
|---|---|
| Principals | Google Account, Service Account, Google Group, Workspace domain, allAuthenticatedUsers, allUsers |
| Roles | Basic (Owner/Editor/Viewer — avoid!), Predefined (service-specific), Custom |
| Policy binding | Grant role to principal on a resource; inherited down the hierarchy |
| Deny policies | Explicitly deny permissions; overrides allow; evaluated before allow |
| IAM Conditions | Conditional bindings (e.g., only during business hours, only for specific resources) |
Service Accounts
- Machine identity; used by applications/VMs
- User-managed: You create; attach to resources
- Default: Auto-created (e.g., GCE default SA); has Editor role — overly permissive, restrict immediately
- Service account impersonation: User acts as SA (requires iam.serviceAccountTokenCreator role)
- Workload Identity (GKE): Maps Kubernetes SA → Google SA; best practice for GKE workloads
- Workload Identity Federation: External identities (AWS, Azure, OIDC) → Google IAM without service account keys; eliminates key management
Key Edge Cases — IAM
- Principle of least privilege: Always the right answer on the exam
- allUsers = anyone on the internet (including unauthenticated)
- allAuthenticatedUsers = anyone with a Google account (still broadly public!)
- Granting Owner at org level = nuclear option — never the right exam answer
- Custom roles can be created at org or project level; can include only needed permissions
- Domain-restricted sharing org policy prevents adding external users to IAM
4.2 Encryption
| Type | Details |
|---|---|
| Default encryption | All data encrypted at rest with Google-managed keys (AES-256); automatic, no config |
| CMEK (Customer-Managed Encryption Keys) | You create/manage keys in Cloud KMS; you control rotation, enable/disable; Google still does the encryption |
| CSEK (Customer-Supplied Encryption Keys) | You provide the key; Google doesn’t store it; only for GCE disks + GCS; you LOSE DATA if you lose the key |
| Client-side encryption | You encrypt before sending to GCP; full control; Google never sees plaintext |
| In-transit encryption | Google encrypts between services (ALTS); HTTPS/TLS for external |
Cloud KMS & Key Management
- Software keys: Default, multi-region capable
- HSM keys (Cloud HSM): Hardware Security Module, FIPS 140-2 Level 3
- External keys (Cloud EKM): Keys stored outside GCP (your HSM); maximum control, higher complexity
- Key rotation: Set automatic rotation schedule; old key versions retained for decryption
Decision: “Customer must control keys but not manage infrastructure” → CMEK. “Customer must hold keys outside GCP” → Cloud EKM. “Regulatory FIPS 140-2 Level 3” → Cloud HSM.
4.3 Network Security
| Service | Purpose |
|---|---|
| VPC Service Controls | Data exfiltration prevention; creates security perimeters around GCP services; blocks data from leaving the perimeter |
| Cloud Armor | WAF + DDoS; attach to External Application LB; preconfigured rules (SQL injection, XSS, OWASP) |
| Identity-Aware Proxy (IAP) | Context-aware access to apps without VPN; verifies identity + device context |
| BeyondCorp Enterprise | Zero-trust access model; combines IAP + Endpoint Verification + Access Context Manager |
| Binary Authorization | Enforce only signed/attested container images can deploy to GKE/Cloud Run |
| Access Context Manager | Define access levels based on IP, device, user attributes; used with VPC SC and IAP |
| Certificate Authority Service | Managed private CA for mTLS, internal certificates |
VPC Service Controls — Critical for Exam
- Creates service perimeters around Google API services (BigQuery, GCS, etc.)
- Blocks data copy/export across perimeter boundaries
- Access levels (via Access Context Manager) define who can cross the perimeter
- Bridge perimeters allow controlled sharing between perimeters
- Dry-run mode: Test policies before enforcing
Edge Case: VPC SC does NOT protect against a malicious insider who already has access within the perimeter. It protects against data leaving the perimeter boundary.
4.4 Secrets & Data Protection
| Service | Use Case |
|---|---|
| Secret Manager | API keys, passwords, certificates; versioned, IAM-controlled, auto-rotation |
| Sensitive Data Protection (Cloud DLP) | Discover, classify, de-identify PII; inspection + transformation |
| Data Loss Prevention | Same as above (rebranded to Sensitive Data Protection) |
| Certificate Manager | Manage TLS certificates for load balancers |
4.5 Security Operations
| Service | Purpose |
|---|---|
| Security Command Center (SCC) | Centralized security dashboard; threat detection, vulnerability scanning, compliance monitoring |
| SCC Premium | Adds Event Threat Detection, Container Threat Detection, VM Threat Detection, Security Health Analytics |
| Chronicle | SIEM platform; ingest and analyze security telemetry at scale |
| Assured Workloads | Compliance controls for regulated workloads (FedRAMP, HIPAA, etc.) |
Compliance Shortcuts
| Requirement | Solution |
|---|---|
| HIPAA | BAA with Google + compliant services + SCC + audit logging |
| PCI DSS | Isolation + VPC SC + CMEK + SCC + encryption everywhere |
| Data residency | Org policy gcp.resourceLocations + regional services |
| FIPS 140-2 | Cloud HSM keys |
| SOC 2 / ISO 27001 | Google already certified; you ensure your configs follow shared responsibility model |
Domain 3 — Practice Questions
Q9. A healthcare company must ensure that data stored in BigQuery cannot be copied to any project outside the organization, even by users with BigQuery Data Viewer access. Which solution should you implement?
A. Remove BigQuery Data Viewer role from all users B. Configure VPC Service Controls with a perimeter around BigQuery C. Enable Cloud Armor on the BigQuery endpoint D. Use CSEK for all BigQuery datasets
Answer: B Why B: VPC Service Controls create a perimeter that prevents data exfiltration across project and org boundaries. Why not A: Removes legitimate access entirely. Why not C: Cloud Armor is for HTTP load balancers, not BigQuery. Why not D: CSEK doesn’t control data copy—only encryption.
Q10. Your GKE workloads need to access Cloud Storage and BigQuery. Currently, all pods use the GCE default service account. A security audit flagged this as a risk. What is the recommended fix?
A. Create a separate service account key for each pod and mount as a Kubernetes secret B. Enable Workload Identity and map Kubernetes service accounts to least-privilege Google service accounts C. Use node-level service accounts with broader permissions D. Grant each namespace Owner role on the project
Answer: B Why B: Workload Identity is the recommended way for GKE pods to authenticate — no keys to manage, least privilege per workload. Why not A: Service account keys are a security anti-pattern (can be leaked). Why not C: Broader permissions = more risk. Why not D: Owner role is wildly overprivileged.
Q11. An organization requires that only container images signed by their CI/CD pipeline can be deployed to production GKE clusters. How should this be enforced?
A. Use an admission controller webhook that checks image tags B. Enable Binary Authorization with an attestor linked to Cloud Build C. Restrict the container registry with IAM to read-only D. Use Network Policy to block unsigned image pulls
Answer: B Why B: Binary Authorization enforces deploy-time attestation — only images with a valid cryptographic attestation from your CI/CD can deploy. Why not A: Image tags are mutable and can be spoofed. Why not C: IAM on the registry doesn’t verify image integrity. Why not D: Network Policy controls network, not image trust.
Q12. A company uses federated identities from their on-premises Active Directory. They need employees to use their existing corporate credentials to access the Google Cloud Console and APIs, without creating Google accounts. What is the best approach?
A. Create Google accounts for every employee manually B. Set up Google Cloud Directory Sync (GCDS) to provision Cloud Identity + SAML SSO with their IdP C. Share a single service account key with all employees D. Use allAuthenticatedUsers on all projects
Answer: B Why B: GCDS syncs AD users/groups to Cloud Identity, and SAML SSO enables login with corporate credentials. Why not A: Manual creation doesn’t scale and isn’t federated. Why not C: Shared credentials = zero accountability. Why not D: Grants access to anyone with any Google account.
5. DOMAIN 4 — ANALYZE & OPTIMIZE TECHNICAL & BUSINESS PROCESSES
5.1 Cost Optimization
Cost Reduction Strategies (by priority)
- Right-size resources: Use Recommender; custom machine types; match CPU/RAM to actual utilization
- Use commitments: CUDs for steady workloads (1yr = ~37% off, 3yr = ~55% off)
- Use Spot/Preemptible VMs: For batch, ML training, CI/CD workers
- Ephemeral resources: Dataproc cluster only during jobs; Cloud Run scale-to-zero
- Storage tiering: Lifecycle policies for GCS; Nearline/Coldline/Archive for infrequent access
- Autoscaling: GCE MIG autoscaler, GKE HPA/VPA, Cloud Run automatic scaling
- BigQuery optimization: Partition + cluster tables; use on-demand for sporadic queries, editions for heavy
BigQuery Pricing Models
| Model | Best For | How It Works |
|---|---|---|
| On-demand | Sporadic/unpredictable queries | Pay per TB scanned; $6.25/TB |
| Editions (Standard/Enterprise/Enterprise Plus) | Heavy, predictable usage | Purchase slots (compute capacity); autoscaling or commitment |
| Flat-rate (legacy) | Being replaced by Editions | Fixed slot allocation |
Tip: “Company runs expensive ad-hoc queries occasionally” → On-demand. “Company runs 100+ scheduled queries daily” → Editions with committed slots.
Compute Pricing Comparison
| Option | Savings | Commitment | Flexibility |
|---|---|---|---|
| On-demand | 0% | None | Full |
| SUDs | Up to 30% | None (automatic) | Full |
| 1yr CUD | ~37% | 1 year | Moderate (resource locked) |
| 3yr CUD | ~55% | 3 years | Low |
| Spot/Preemptible | Up to 91% | None | Workload must be fault-tolerant |
5.2 Performance Optimization
| Technique | When / How |
|---|---|
| Caching | Memorystore (Redis) for app sessions; Cloud CDN for static content; BigQuery BI Engine for dashboards |
| Read replicas | Cloud SQL / AlloyDB for read-heavy workloads |
| Connection pooling | Cloud SQL Proxy + pgbouncer for high-connection apps |
| Global load balancing | Serve users from nearest region via anycast IP |
| GKE HPA/VPA | HPA: scale pods horizontally by CPU/custom metrics. VPA: resize pods vertically. Don’t use both on same metric |
| Cloud CDN | Reduce latency + backend load for static/semi-static content |
| Storage class | Standard GCS for frequently accessed; Local SSD for ultra-low latency temporary |
| BigQuery optimization | Partition, cluster, use approximate functions, avoid SELECT * |
5.3 Monitoring & Observability
| Service | Purpose |
|---|---|
| Cloud Monitoring | Metrics, dashboards, alerting policies, uptime checks, SLOs |
| Cloud Logging | Centralized log ingestion, analysis, export; log-based metrics; log sinks |
| Cloud Trace | Distributed tracing for latency analysis in microservices |
| Cloud Profiler | Continuous CPU/memory profiling for production code |
| Error Reporting | Aggregates and tracks application errors; alerts on new errors |
Logging Architecture
- Log Sinks: Route logs to BigQuery (analysis), GCS (archive), Pub/Sub (streaming), Splunk
- Log Router: Processes all logs; inclusion/exclusion filters
- Exclusion Filters: Reduce log volume (and cost) by excluding verbose logs
- Retention: Default 30 days (_Default bucket); custom buckets up to 10 years
- Log-based Metrics: Create custom metrics from log entries for alerting/dashboards
Alerting Best Practices
- Use multi-condition policies (e.g., high latency AND high error rate)
- Set notification channels (email, Slack, PagerDuty, Pub/Sub)
- Define SLIs (request latency, error rate, throughput) and SLOs (99.9% availability)
- Use error budgets to balance reliability and feature velocity
Domain 4 — Practice Questions
Q13. A data analytics team runs BigQuery queries that cost $50,000/month. Usage is predictable with 200+ scheduled queries daily. Which pricing change would most reduce costs?
A. Switch to BigQuery on-demand pricing B. Purchase BigQuery Editions with committed slots C. Move all data to Cloud SQL to avoid BigQuery costs D. Use Cloud Storage to store query results and re-read them
Answer: B Why B: Heavy, predictable usage benefits from committed slots → lower per-query cost. Why not A: They likely already use on-demand at this spend; switching to it wouldn’t help. Why not C: Cloud SQL isn’t for analytics at this scale. Why not D: Doesn’t reduce query cost.
Q14. Your GKE application experiences latency spikes during traffic surges. Pods take 2 minutes to start. How can you reduce the impact? (Select TWO)
A. Configure Horizontal Pod Autoscaler (HPA) with custom metrics and lower target utilization B. Use Vertical Pod Autoscaler (VPA) to give each pod more resources C. Enable GKE node auto-provisioning to add nodes faster D. Switch to Bare Metal Solution for lower latency
Answer: A, C Why A: HPA with lower target triggers scaling earlier, before load spikes. Why C: Node auto-provisioning ensures node capacity is available for new pods. Why not B: VPA resizes pods but doesn’t add more pods for traffic spikes. Why not D: BMS doesn’t solve autoscaling latency.
Q15. Your team needs to detect anomalous login attempts to Google Cloud Console and receive real-time alerts. What should you implement?
A. Cloud Monitoring uptime checks on the Console URL B. Export Admin Activity audit logs to BigQuery and run scheduled queries C. Enable Security Command Center Event Threat Detection + alert via Pub/Sub D. Configure Cloud CDN to block suspicious IPs
Answer: C Why C: SCC Event Threat Detection analyzes audit logs in real time for threats like suspicious logins and brute force. Why not A: Uptime checks verify availability, not security. Why not B: Scheduled queries are not real-time. Why not D: CDN doesn’t protect Console access.
6. DOMAIN 5 — MANAGE IMPLEMENTATIONS OF CLOUD ARCHITECTURE
6.1 Migration — The 6 Rs
| Strategy | Description | GCP Service |
|---|---|---|
| Rehost (Lift & Shift) | Move VM as-is | Migrate to VMs |
| Replatform | Minor changes (e.g., managed DB) | Migrate to VMs → Cloud SQL |
| Refactor | Re-architect for cloud native | Rewrite for GKE / Cloud Run |
| Repurchase | Move to SaaS | Switch to Google Workspace, etc. |
| Retire | Decommission | N/A |
| Retain | Keep on-prem | Hybrid with Interconnect |
Migration Tools
| Tool | Use Case |
|---|---|
| Migrate to VMs | Migrate VMs from VMware, AWS, Azure, physical to GCE |
| Migrate to Containers | Convert VMs to containers for GKE |
| Database Migration Service (DMS) | Migrate MySQL, PostgreSQL, SQL Server, Oracle → Cloud SQL / AlloyDB |
| BigQuery Data Transfer Service | Schedule transfers from SaaS (Google Ads), GCS, S3 → BigQuery |
| Transfer Appliance | Offline data transfer for 100+ TB; physically ship a device |
| Storage Transfer Service | Online transfer from S3, Azure Blob, HTTP, other GCS buckets |
| gcloud storage / gsutil | Command-line data transfer; for smaller datasets or scripted transfers |
| Datastream | Continuous CDC replication from Oracle/MySQL |
| VMware Engine | Run VMware workloads natively on GCP; minimal changes |
Migration Approach — Phases
- Assess: Inventory workloads; classify by 6 Rs; identify dependencies (StratoZone, Migration Center)
- Plan: Design target architecture; sizing; networking; identity; compliance requirements
- Deploy foundation: Landing zone, VPC, IAM, billing, logging
- Migrate: Wave-based approach; start with non-critical workloads; verify each wave
- Optimize: Right-size, apply discounts, modernize incrementally
6.2 CI/CD & DevOps
| Service | Purpose |
|---|---|
| Cloud Build | Managed CI/CD; builds containers, runs tests, deploys; serverless |
| Cloud Deploy | Managed continuous delivery to GKE/Cloud Run/GCE; promotion through environments |
| Artifact Registry | Store containers, packages (Maven, npm, Python); vulnerability scanning |
| Cloud Source Repositories | Git repos on GCP (or mirror from GitHub/Bitbucket) |
| Container Registry (deprecated) | → Migrate to Artifact Registry |
Cloud Build Key Features
- Build triggers: Commit, tag, PR on GitHub/Bitbucket/CSR
- Private pools: Run builds in your VPC for accessing private resources
- Approval gates (Cloud Deploy): Require manual approval before promoting to production
- Build steps: Custom containers for each step; parallel/sequential execution
GKE Deployment Strategies
| Strategy | How It Works | Risk Level |
|---|---|---|
| Rolling update | Gradually replace pods | Low; default |
| Blue-green | Two full environments; switch traffic | Lowest risk; highest cost |
| Canary | Small % traffic to new version | Low; validates in production |
| A/B testing | Route by headers/cookies | Application-level; for experiments |
6.3 GKE Deep Dive
Key Concepts
| Concept | Details |
|---|---|
| Autopilot vs Standard | Autopilot: Google manages nodes, pod-level billing, secure defaults. Standard: you manage node pools, more control |
| Node pools | Groups of VMs with same config; different pools for different workload types |
| Private clusters | Nodes have internal IPs only; control plane can be private or with authorized networks |
| Release channels | Rapid, Regular, Stable; controls auto-upgrade cadence |
| Network Policy | Kubernetes-level firewall; control pod-to-pod traffic |
| Gateway API | Next-gen ingress; multi-cluster, traffic splitting, header routing |
| Config Sync | GitOps for GKE; sync cluster config from Git repo |
| Policy Controller | Enforce policies on Kubernetes resources (OPA Gatekeeper) |
GKE Enterprise (formerly Anthos)
- Multi-cluster management across GCP, on-prem, other clouds (AWS, Azure)
- Service Mesh: Managed Istio for mTLS, traffic management, observability
- Config Sync + Policy Controller for GitOps and policy enforcement across fleets
- Connect Gateway: Register external clusters to GCP for centralized management
- Fleet: Logical grouping of GKE clusters for unified management
Domain 5 — Practice Questions
Q16. A company is migrating 200 VMs from their VMware on-premises data center to GCP. They want minimal changes and need to maintain their existing VM images. What is the most efficient migration approach?
A. Rebuild all applications as containers on GKE B. Use Migrate to VMs for automated rehost to Compute Engine C. Manually create disk images and upload via gcloud D. Use Transfer Appliance to ship VM disks physically
Answer: B Why B: Migrate to VMs automates the rehost of VMware VMs to Compute Engine with minimal changes. Why not A: “Minimal changes” rules out containerization. Why not C: Manual approach doesn’t scale for 200 VMs. Why not D: Transfer Appliance is for data, not VM migration.
Q17. Your team uses Cloud Build for CI/CD. You need to ensure that the build process can pull private npm packages from a registry inside your corporate VPC. What should you configure?
A. Open the corporate firewall to allow Cloud Build’s public IP range B. Use Cloud Build private pools with VPC peering to your network C. Store npm packages in Cloud Storage and download via signed URLs D. Embed npm credentials in the Dockerfile
Answer: B Why B: Private pools run inside your VPC, enabling access to private resources without exposing them publicly. Why not A: Opening firewalls to public IPs is a security anti-pattern. Why not C: Not a standard npm workflow; fragile. Why not D: Embedding credentials in Dockerfiles is a security vulnerability.
Q18. A gaming company needs zero-downtime deployments for their GKE-based game backend. They want to test new versions with 5% of live traffic before full rollout. Which deployment strategy should they use?
A. Rolling update with maxUnavailable: 0 B. Blue-green deployment with manual traffic switch C. Canary deployment using Cloud Deploy with traffic splitting D. Recreate deployment (terminate all, then deploy new)
Answer: C Why C: Canary with traffic splitting sends exactly 5% of traffic to the new version for validation. Why not A: Rolling update replaces pods gradually but doesn’t give 5% traffic control. Why not B: Blue-green is all-or-nothing traffic switch. Why not D: Recreate causes downtime.
7. DOMAIN 6 — ENSURE SOLUTION & OPERATIONS EXCELLENCE
7.1 Reliability & Disaster Recovery
Availability & SLA Concepts
| Term | Meaning |
|---|---|
| SLA | Service Level Agreement — contractual availability guarantee |
| SLO | Service Level Objective — internal target (stricter than SLA) |
| SLI | Service Level Indicator — measured metric (latency, error rate, uptime) |
| Error Budget | 100% − SLO = budget for risk (deployments, experiments) |
| RPO | Recovery Point Objective — max acceptable data loss (time) |
| RTO | Recovery Time Objective — max acceptable downtime |
DR Patterns
| Pattern | RPO | RTO | Cost | Description |
|---|---|---|---|---|
| Cold standby | Hours | Hours | Lowest | Backups only; restore infrastructure when needed |
| Warm standby / Pilot light | Minutes | Minutes-Hours | Medium | Minimal infrastructure running; scale up on failover |
| Hot standby / Active-active | Near-zero | Seconds-Minutes | Highest | Full environment in 2+ regions; automatic failover |
GCP HA / DR Features by Service
| Service | HA Feature |
|---|---|
| Cloud SQL | Multi-zone HA (automatic failover); cross-region read replicas; PITR |
| Cloud Spanner | Multi-region configs (99.999% SLA); automatic replication |
| GKE | Multi-zone clusters; regional clusters (control plane in 3 zones) |
| Cloud Storage | Multi-region / dual-region buckets; 11 9s durability |
| GCE | Regional MIGs; autohealing; live migration (default for maintenance) |
| Cloud Run | Multi-region traffic splitting via External LB |
| AlloyDB | Cross-region replication; read pool for HA |
Backup Strategies
- Cloud SQL: Automated backups + on-demand backups; PITR with binary logging
- GCE: Scheduled disk snapshots (incremental); machine images for full VM backup
- GCS: Object versioning; bucket lock for retention compliance
- GKE: Backup for GKE (managed backup/restore of cluster resources + PVs)
7.2 Operations
Cloud Operations Suite Cheat Sheet
| Tool | Use Case | Remember |
|---|---|---|
| Cloud Monitoring | Metrics, SLOs, alerting, uptime checks | Set up alerting policies with multiple notification channels |
| Cloud Logging | Centralized logs, search, analysis | Log sinks export to BigQuery (analysis), GCS (archive), Pub/Sub (streaming) |
| Cloud Trace | Latency profiling (distributed tracing) | Find bottlenecks in microservices |
| Cloud Profiler | CPU/memory profiling in production | Minimal overhead; continuous |
| Error Reporting | Group and track errors | Auto-notification for new errors |
Audit Logging
| Log Type | What It Captures | Always On? |
|---|---|---|
| Admin Activity | Resource creation, deletion, IAM changes | Yes (free, cannot disable) |
| Data Access | Read/write to resource data (BigQuery queries, GCS reads) | No (must enable; can be expensive) |
| System Event | Google-initiated actions (live migration, auto-scaling) | Yes (free) |
| Policy Denied | Access denied by VPC SC, org policies | Yes (free) |
Domain 6 — Practice Questions
Q19. Your company requires an RPO of 5 minutes and RTO of 30 minutes for their Cloud SQL production database. A regional outage occurs. What should your DR architecture include?
A. Automated backups with daily frequency B. Cross-region read replica with automated promotion C. Multi-zone HA only (no cross-region) D. Export database to Cloud Storage daily
Answer: B Why B: Cross-region read replica with promotion meets RPO ~minutes (async replication) and RTO ~30 min (promotion time). Why not A: Daily backups = RPO of up to 24 hours. Why not C: Multi-zone HA protects against zone failures, not regional outages. Why not D: Daily exports = RPO 24 hours.
Q20. A production microservices application on GKE experiences intermittent latency spikes. Users report slow responses. Which tool should you use first to identify the root cause?
A. Cloud Logging to search for error messages B. Cloud Trace to analyze distributed request latency C. Cloud Profiler to check CPU usage D. Cloud Monitoring to check VM memory
Answer: B Why B: Cloud Trace shows end-to-end request latency across microservices, identifying which service is the bottleneck. Why not A: Logging might not show latency issues if there are no errors. Why not C: Profiler shows code-level CPU/memory but not request flow. Why not D: VM metrics don’t show per-request latency in a microservices context.
Q21. An audit requires that all BigQuery query logs from the past 2 years are stored in immutable storage for compliance. How should you configure this?
A. Enable Data Access audit logs for BigQuery; create a log sink to a GCS bucket with a retention policy and Bucket Lock B. Export BigQuery job metadata using a Cloud Function daily C. Store audit logs in the default _Default log bucket D. Enable Cloud CDN logging for BigQuery
Answer: A Why A: Data Access audit logs capture query activity; sinking to GCS with Bucket Lock ensures immutable, long-term retention. Why not B: Custom export is fragile and misses real-time entries. Why not C: Default bucket retention is 30 days, not 2 years. Why not D: CDN logging is irrelevant to BigQuery.
8. WELL-ARCHITECTED FRAMEWORK
6 Pillars Summary
| Pillar | Focus | Key Question on Exam |
|---|---|---|
| Operational Excellence | Monitoring, incident response, automation, toil reduction | “How do you ensure ops visibility?” |
| Security, Privacy, Compliance | IAM, encryption, audit, VPC SC, compliance frameworks | “How do you protect data and meet regulations?” |
| Reliability | HA, DR, fault tolerance, chaos engineering, SLOs | “What happens when a region fails?” |
| Cost Optimization | Right-sizing, CUDs, storage tiering, recommender | “How do you minimize cost meeting requirements?” |
| Performance Optimization | Caching, autoscaling, CDN, profiling | “How do you reduce latency?” |
| Sustainability | Carbon-aware scheduling, region selection, efficient resources | “How do you reduce environmental impact?” |
Core Principles (memorize these)
- Design for change — DORA metrics, CI/CD, small frequent deploys
- Document your architecture — Shared understanding, decision history
- Simplify and use managed services — Reduce operational burden
- Decouple your architecture — Independent scaling, upgrades, security
- Use stateless architecture — Easier scaling, resilience, restart
9. CASE STUDIES DEEP DIVE
2 of these 4 will appear on your exam. Each is shown on a split-screen tab. Skim before answering. Focus on business requirements, technical requirements, and existing environment.
9.1 EHR Healthcare
Company: Electronic Health Records provider Key Requirements:
- HIPAA compliance (BAA, encryption, audit logging)
- Migrate on-prem data centers to GCP
- Real-time data processing for patient records
- Scalable APIs for partner hospitals
- Disaster recovery with RPO < 1h, RTO < 4h
- Machine learning for predictive analytics
Architecture Pointers:
- Cloud Healthcare API (FHIR, HL7, DICOM) for medical data standards
- Cloud SQL / AlloyDB for relational patient data with HA
- BigQuery for analytics + BigQuery ML for predictions
- Pub/Sub + Dataflow for real-time data pipelines
- VPC Service Controls to prevent data exfiltration
- CMEK for encryption (HIPAA requirement)
- Identity-Aware Proxy for secure access
- SCC Premium for threat detection
- DR: Cross-region replicas + regular backups
9.2 Helicopter Racing League (HRL)
Company: Helicopter racing entertainment company Key Requirements:
- Real-time telemetry from helicopters (IoT-scale data ingestion)
- Live video streaming to fans worldwide
- Real-time predictions and analytics during races
- Predictive maintenance for helicopter fleet
- Low-latency global content delivery
- Cost-effective storage for historical race data
Architecture Pointers:
- Pub/Sub for real-time telemetry ingestion
- Dataflow for stream processing of telemetry
- Cloud Bigtable for time-series telemetry storage (low-latency, massive scale)
- BigQuery for historical analytics
- Vertex AI for predictive maintenance models
- Cloud CDN / Media CDN for global video delivery
- Cloud Storage with lifecycle policies (Standard → Nearline → Coldline) for historical video
- GKE for backend microservices
- External Application LB + Cloud Armor for global distribution + security
9.3 TerramEarth
Company: Heavy equipment manufacturer (mining/construction vehicles) Key Requirements:
- IoT sensors on 500,000+ vehicles sending telemetry
- Limited vehicle connectivity (batch uploads when near base stations)
- Predictive maintenance to prevent equipment failure
- Dealer network APIs for parts ordering
- 2 PB total data, growing 200 TB/year
- Data analytics for fleet optimization
Architecture Pointers:
- Edge processing on vehicles (limited connectivity)
- Cloud Storage for batch uploads of telemetry data
- Pub/Sub for near-real-time data when connected
- Dataflow for processing
- Bigtable for time-series sensor data
- BigQuery for fleet analytics
- Vertex AI for predictive maintenance
- Apigee for dealer/partner API management
- Dedicated Interconnect for high-bandwidth data center connection
- Storage lifecycle management for 2 PB+ data
9.4 Mountkirk Games
Company: Online mobile gaming Key Requirements:
- Backend for multiplayer mobile game
- Massive scale (millions of concurrent users)
- Global deployment for low latency
- Real-time leaderboards
- Analytics on player behavior for game design decisions
- GPU for graphics rendering pipeline
- Cost optimization during off-peak
Architecture Pointers:
- GKE (Autopilot or Standard with Spot nodes) for game backend; multi-region
- Cloud Spanner for global leaderboard + player state (strong consistency, global)
- Memorystore (Redis) for session caching and leaderboard cache
- Pub/Sub + Dataflow → BigQuery for player analytics pipeline
- External Application LB with Cloud CDN for global distribution
- Spot VMs / preemptible nodes for batch processing (off-peak cost savings)
- GCE with GPUs for rendering pipeline
- Cloud Monitoring with SLOs for game reliability
Case Studies — Practice Questions
Q22. (EHR Healthcare) The EHR team needs to process HL7 FHIR patient records from 50 partner hospitals in real time and must maintain HIPAA compliance. Which pipeline architecture is most appropriate?
A. Partners upload files to Cloud Storage → trigger Cloud Function → insert into Cloud SQL B. Partners send FHIR data to Cloud Healthcare API → Pub/Sub notifications → Dataflow processing → BigQuery analytics C. Partners connect via FTP to a Compute Engine server → manual import to BigQuery D. Partners email CSV files → Cloud Functions parses emails
Answer: B Why B: Cloud Healthcare API natively handles FHIR format; Pub/Sub + Dataflow provides real-time processing; BigQuery for analytics. All HIPAA-compliant. Why not A: Cloud Storage batch uploads aren’t real-time. Why not C/D: FTP and email are not scalable or compliant approaches.
Q23. (Mountkirk Games) The live game backend experiences 10 million concurrent users during peak events. Leaderboard data must be globally consistent. Which database handles the leaderboard?
A. Cloud SQL with read replicas B. Firestore in Native mode C. Cloud Spanner (multi-region) D. Cloud Bigtable
Answer: C Why C: Spanner provides global strong consistency and horizontal scaling for 10M concurrent users. Why not A: Cloud SQL doesn’t scale horizontally for this volume. Why not B: Firestore lacks strong consistency for global ranked leaderboards. Why not D: Bigtable is eventually consistent and lacks SQL for ranking queries.
Q24. (TerramEarth) Vehicles upload sensor data in batches when near base stations. The data must be processed and made available for predictive maintenance models within 2 hours. The most cost-effective architecture is:
A. IoT Core → Pub/Sub → Dataflow → BigQuery (real-time streaming) B. Batch upload to Cloud Storage → Cloud Function trigger → Dataflow batch → BigQuery → Vertex AI C. FTP to on-prem server → manual upload to BigQuery D. Direct database writes from each vehicle to Cloud Spanner
Answer: B Why B: Batch upload matches the intermittent connectivity; GCS trigger → Dataflow batch is cost-effective; 2-hour SLA doesn’t require real-time. Why not A: IoT Core is for real-time connected devices; vehicles have limited connectivity. Why not C: On-prem server defeats cloud migration purpose. Why not D: Direct writes require constant connectivity.
Q25. (Helicopter Racing League) The HRL needs to deliver live race video globally with sub-100ms latency and protect the stream from DDoS attacks. Which services should they combine? (Select TWO)
A. Media CDN for global video delivery B. Cloud SQL for video storage C. Cloud Armor with rate limiting and WAF rules D. Cloud Functions to transcode video in real-time
Answer: A, C Why A: Media CDN is optimized for low-latency media delivery at global scale. Why C: Cloud Armor provides DDoS protection and WAF rules to protect the delivery infrastructure. Why not B: SQL databases don’t serve video. Why not D: Cloud Functions isn’t for real-time video transcoding at scale.
10. DECISION TREES & CHEAT SHEETS
10.1 Which Compute?
Need custom OS / GPU / BYOL / legacy? → Compute Engine
Need containers + orchestration + hybrid? → GKE
Need stateless HTTP container, low traffic? → Cloud Run
Need single-purpose event trigger? → Cloud Functions
Need simple web app, auto-scale to zero? → App Engine Standard
Need custom runtime, WebSockets? → App Engine Flexible
Need batch HPC jobs? → Cloud Batch
Need VMware workloads on GCP? → VMware Engine
Need Oracle DB / specialized hardware? → Bare Metal Solution
10.2 Which Database?
Relational + global + 99.999%? → Cloud Spanner
Relational + PostgreSQL + analytics? → AlloyDB
Relational + MySQL/PG/MSSQL + simple? → Cloud SQL
NoSQL + key-value + massive + IoT? → Cloud Bigtable
NoSQL + document + mobile sync? → Firestore (native mode)
Cache + sessions + leaderboard? → Memorystore (Redis)
Analytics warehouse? → BigQuery
10.3 Which Storage?
Objects (files, images, backups)? → Cloud Storage
VM disks? → Persistent Disk
Shared file system (NFS)? → Filestore
Ultra-low latency ephemeral? → Local SSD
10.4 Which Migration Tool?
VMs from VMware/AWS/Azure? → Migrate to VMs
VMs to containers? → Migrate to Containers
MySQL/PG/SQL Server → Cloud SQL? → Database Migration Service
100+ TB offline? → Transfer Appliance
S3/Azure Blob → GCS? → Storage Transfer Service
VMware workloads as-is? → VMware Engine
Oracle/MySQL CDC? → Datastream
10.5 Which Interconnect?
High bandwidth + colocation available? → Dedicated Interconnect
High bandwidth + no colocation? → Partner Interconnect
Encrypted tunnel + small bandwidth? → HA VPN
GCP ↔ AWS/Azure direct? → Cross-Cloud Interconnect
Offline bulk data transfer? → Transfer Appliance
10.6 Which Security Service?
Stop data exfiltration from GCP APIs? → VPC Service Controls
WAF + DDoS for web apps? → Cloud Armor
Context-aware app access without VPN? → Identity-Aware Proxy (IAP)
Zero-trust complete model? → BeyondCorp Enterprise
Only signed containers can deploy? → Binary Authorization
Discover PII in data? → Sensitive Data Protection
Store API keys and secrets? → Secret Manager
Encrypt with your keys (managed)? → CMEK (Cloud KMS)
Encrypt with keys outside GCP? → Cloud EKM
FIPS 140-2 Level 3? → Cloud HSM
Centralized security posture? → Security Command Center
SIEM at scale? → Chronicle
Compliance guardrails? → Assured Workloads
External identity federation? → Workload Identity Federation
10.7 Which API Management?
Full lifecycle API management? → Apigee
Simple API gateway? → API Gateway
gRPC API routing? → Cloud Endpoints
10.8 Which Container Registry?
New projects? → Artifact Registry
Legacy? → Container Registry (deprecated)
11. CRITICAL EDGE CASES & EXAM TRAPS
Networking Traps
- VPC Peering is NOT transitive. A ↔ B and B ↔ C does NOT give A ↔ C
- Shared VPC IS transitive through the host project
- Private Google Access must be enabled per subnet for private VMs to reach Google APIs
- Cloud NAT does NOT provide inbound access — it’s outbound only
- HA VPN requires 2 tunnels (one per interface) for 99.99% SLA
- Dedicated Interconnect requires 2 connections in different edge availability domains for 99.99% SLA
- Cloud Armor only works with External Application Load Balancer (not internal LBs or network LBs)
IAM Traps
- Default service accounts have Editor role — always a wrong answer when security is mentioned
- allAuthenticatedUsers includes ANYONE with a Google account — this is NOT your org only
- Org policies CANNOT be overridden by child nodes; they only become more restrictive
- Deny policies take precedence over allow policies
- Service account keys should be avoided — use Workload Identity or Workload Identity Federation instead
- Custom roles can be created at project or org level, not folder level
Storage Traps
- Object Lifecycle can only transition to COLDER classes (Standard → Nearline → Coldline → Archive), never warmer
- Bucket Lock is IRREVERSIBLE — once set, the retention policy cannot be removed
- Firestore mode (Native vs Datastore) cannot be changed after database creation
- Bigtable performs poorly with small datasets (<1 TB); designed for massive scale
- Regional Persistent Disk provides HA across zones but with 2x cost
Database Traps
- Cloud SQL max storage: 64 TB. For larger, consider Spanner or AlloyDB
- Cloud Spanner: Minimum cost is ~$0.90/hr per node. Not cost-effective for small workloads
- AlloyDB: PostgreSQL compatible but NOT open-source — it’s Google proprietary
- BigQuery: Querying with SELECT * on partitioned table still scans all partitions unless you filter on partition column
GKE Traps
- Autopilot doesn’t allow SSH to nodes, privileged containers, or host network access
- HPA and VPA on the same metric will conflict — don’t use both on CPU simultaneously
- Workload Identity requires a 1:1 mapping between K8s SA and Google SA
- Node auto-repair and node auto-upgrade are independent settings
Cost Traps
- Sustained Use Discounts are automatic but only apply to GCE and GKE Standard nodes (NOT Autopilot)
- Committed Use Discounts cannot be cancelled; you pay for the full term
- Preemptible/Spot VMs can be reclaimed with 30-second notice — NOT suitable for stateful workloads
- Cloud Run min-instances > 0 means you pay even with no traffic (but eliminates cold starts)
Data Traps
- Pub/Sub guarantees at-least-once delivery — your consumer must handle duplicates
- Dataflow (Apache Beam) can do both batch AND stream — it’s not stream-only
- Dataproc is for existing Spark/Hadoop jobs — don’t pick it for new greenfield ETL (use Dataflow)
- Cloud Composer orchestrates pipelines — it does NOT process data itself
Security Traps
- VPC Service Controls protect API-level access, NOT network-level (that’s firewall rules / Cloud Armor)
- CSEK keys are NOT stored by Google — lose the key = lose the data permanently
- Cloud KMS key rotation creates new key VERSION — old versions still decrypt old data
- Binary Authorization requires attestation BEFORE deployment, not after
- Data Access audit logs are disabled by default and can be expensive — explicitly enable them for compliance
Migration Traps
- Migrate to VMs streams disk data during replication — minimal downtime for cutover
- Database Migration Service supports continuous replication (not just one-time)
- Transfer Appliance can take weeks (ship → ingest → ship back) — plan ahead
- VMware Engine keeps your VMware investment — use when “minimal change” AND VMware is mentioned
FINAL EXAM DAY CHECKLIST
- Read each question completely — identify ALL constraints before looking at answers
- Look for absolute words (“always,” “never,” “only”) — these are often wrong
- “Minimize operational overhead” = most managed service
- “Minimize cost” = cheapest that meets ALL requirements (not just cheapest)
- “Most secure” = least privilege + encryption + network isolation
- If two options are technically correct, choose the more Google-native solution
- Case study questions: re-read the relevant requirement section before answering
- Multi-select: count answers carefully (select exactly what’s asked)
- Time management: ~2 min per question; flag and return to hard ones
- Trust your first instinct unless you find a clear reason to change
Last updated: June 2025. Based on the official PCA Standard Exam Guide (2026 version), Well-Architected Framework, and all 4 case studies (EHR Healthcare, Helicopter Racing League, TerramEarth, Mountkirk Games).