cert-bank

Guides

이 글을 한국어로 읽기

GCP-PCA Study Guide: Turn Every Requirement Into a Number, Then Defend the Cheapest Design That Meets It

The Professional Cloud Architect exam rarely asks whether you know a product. It asks whether you can defend a design: this option, not that one, because of a requirement written somewhere in the stem or in a case study. Most wrong options are good designs for a different set of requirements. So this guide is not organised by product. It is organised by the habits that let you find the requirement that decides, put a number on it, and pick the cheapest design that meets it.

Study guide · Google Cloud Certified Professional Cloud Architect (GCP-PCA)

GCP-PCA Study Guide: Turn Every Requirement Into a Number, Then Defend the Cheapest Design That Meets It

The exam

50 to 60 multiple choice and multiple select questions in 120 minutes, so roughly two minutes each. Google does not publish a pass mark; check the official exam page for current details. Six sections: Designing and planning a cloud solution architecture (25%), Managing and provisioning a cloud solution infrastructure (17.5%), Designing for security and compliance (17.5%), Analyzing and optimizing technical and business processes (15%), Managing implementation (12.5%) and Ensuring solution and operations excellence (12.5%).

Google publishes four case studies (Altostrat Media, Cymbal Retail, EHR Healthcare and KnightMotives Automotive); two of them appear on your exam, and together they make up 20 to 30% of the questions. In the 274 GCP-PCA questions in this question bank, 84 (31%) are set inside one of the four cases, and 58 (21%) are trade-off questions, the highest share of any Google Cloud exam in our bank.

Habit 1: turn adjectives into numbers

A functional requirement says what a system does. A non-functional requirement says how well: latency, availability, scalability, security, compliance, cost, operability. Architects spend most of their effort on the second kind, because those decide regions, replicas and spend. The exam rewards the candidate who reads "always available" and asks "how available, for which path?".

When a stem gives a number, take it literally. A 15-minute RPO rules out any nightly backup. A 4-hour RTO leaves time to scale out a prepared stack, so a full hot copy is unnecessary. "At least 99.9% for customer-facing systems" (EHR Healthcare's promise) applies to the whole request path, not to each component. When the stem gives no number and only says "lowest" or "highest", look for another constraint that limits it, usually cost or operational effort.

Going deeper: availability math for a chain of components

When components are called one after another and all must work, their availabilities multiply. A web tier, an API tier and a database at 99.9% each give roughly 99.7% end to end, which is why a company cannot promise customers 99.9% just because every component has that objective. Redundant copies in parallel work the other way: the pair fails only when both copies fail at the same time, so two independent 99.9% copies are far better than one. The key word is independent. Two copies in the same zone, behind the same single load balancer, or sharing one configuration, fail together. That is why the exam keeps asking where the copies live: spread them across zones to survive a zone failure, across regions to survive a regional one. The third lever is to remove hard serial dependencies, for example by making a non-essential call asynchronous.

Habit 2: match the recovery pattern to RTO and RPO, not to its name

PatternWhat exists in the recovery locationFits
ColdBackups only; recovery means provisioning and restoringLong RTO, RPO = age of last backup
Pilot lightData replicated continuously; compute defined as code, started after a disasterShort RPO, moderate RTO, low cost
WarmData replicated, a scaled-down stack runningShort RPO, RTO of hours or less
HotBoth sides serve production trafficSmallest RTO and RPO, highest cost

Google's disaster recovery planning guide names three patterns: cold, warm and hot. Pilot light is a common industry name for the lean end of warm, where only the data tier runs. Choose the cheapest row whose replication meets the RPO and whose recovery steps fit inside the RTO. A logistics company with a 4-hour RTO, a 15-minute RPO and no budget for a second production stack is warm standby: a cross-region replica plus a scaled-down copy of the application.

Then add the rule people forget: redundancy is not backup. Replicas and HA standbys copy every change, including a dropped table or an attacker's encryption, within seconds. Recovering from logical damage needs an earlier state: point-in-time recovery, immutable backups in a backup vault, object versioning, and a restore that has actually been tested.

Which recovery pattern do the RTO and RPO allow?
Is a long RTO acceptable, with RPO equal to the age of the last backup?
YES →
Cold: backups only
↓ no
Short RPO, moderate RTO, low cost, compute started after a disaster?
YES →
Pilot light: data replicated, compute as code
↓ no
Short RPO and an RTO of hours or less?
YES →
Warm: replica plus a scaled-down stack
↓ no
OTHERWISE →
Hot: both sides serve production traffic, at the highest cost
Rows go from cheapest to dearest: the first one that meets both numbers is the design to pick.

Habit 3: read a case study like a requirements document

Read each published case before the exam, and reduce it to the facts that will decide questions. A few examples of what to note:

  • Altostrat Media: a large media library in Cloud Storage, legacy on-premises ingestion, and needs for hybrid connectivity, storage cost with high availability, and explainable, auditable AI with harmful-content detection.
  • Cymbal Retail: MySQL, SQL Server, Redis and MongoDB, SFTP and batch ETL integrations, and gen AI for catalogue enrichment (with human review before catalogue updates) and conversational commerce.
  • EHR Healthcare: leaving colocation, at least 99.9% availability, legacy insurer interfaces, consistent logging and monitoring, compliance.
  • KnightMotives Automotive: a mainframe ERP, past breaches, EU data protection duties, weak plant connectivity and a hybrid strategy.

Answer each case question for that case only. A fact from another case, or a requirement you invented, is the most common way to talk yourself out of the right answer.

Beginner tip: for each of the four case studies, write one card with three columns: hard numbers (availability, RTO, RPO, regions), hard constraints (regulations, legacy systems that must stay, skills of the team) and stated goals. Then, for every case question you practise, circle the single item on the card that decided the answer. After a few sets you will find you read the card, not the question, first.

Habit 4: let the hard constraint eliminate options

Most stems contain one requirement that some options break outright. Underline it, then cross out every option that breaks it before comparing the rest. Typical deciders:

  • "Data may not leave the EU" removes any global or out-of-region design; the resource locations constraint and Assured Workloads are the tools.
  • "Traffic must not use the public internet" removes HA VPN and leaves Interconnect. A 99.99% hybrid design needs Google's 99.99% Dedicated Interconnect topology (at least four connections, spread across both edge availability domains), not one connection with a VPN backup.
  • "Writes from several continents, strongly consistent" removes Cloud SQL and AlloyDB, which send every write to one primary, and leaves Spanner.
  • "Minimal changes to the application" removes refactoring options, however modern.
  • "Lowest operational overhead" favours the managed service when nothing else forbids it.

Habit 5: the process questions

About a quarter of the exam (Analyzing and optimizing processes, Managing implementation) is about the work around the technology. Look for the verb. Delivery: build once, promote the same artifact with Cloud Deploy and approvals. Failure: hypothesis, evidence, mitigate first, root cause after. People: stakeholder analysis, pilots, training. Decisions: reversible ones quickly; hard-to-reverse ones (where regulated data lives) with a proof of concept against agreed criteria and an architecture decision record. Money: attribute costs first, remove waste, then commit only for the steady baseline. The usual wrong answers are a bigger instance, a mandate from above, or a plan that was never tested.

A study order

  1. Read the four case studies once, then build the cards from the tip above.
  2. Design and planning (25%): requirements, availability, DR patterns, compute, storage and database choices.
  3. Networking and hybrid connectivity, then security and compliance: hierarchy, IAM, VPC Service Controls, keys, residency.
  4. Delivery, migration and operations, then mixed sets, flagging long case questions and returning to them.

Practise

Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.

Approximate section weights from Google's Professional Cloud Architect exam guide; question count, time and case-study rules from the official exam page. The case studies are revised from time to time; read the current versions linked from the exam guide.