cert-bank

Guides

이 글을 한국어로 읽기

9 Question Patterns People Get Wrong Across the Google Cloud Exams, and the Rule Behind Each

The six Google Cloud exams look different on paper: two are for business leaders, one for engineers who run projects day to day, and the others for architects, security engineers and data engineers. But the same traps come back in all of them, dressed in different products. We read the "key idea" line of every explanation in the 1,601 Google Cloud questions we wrote for this question bank and grouped the ones where a plausible option fails for the same underlying reason. These are the nine patterns that recur across exams, with the rule that settles each.

Common mistakes · Google Cloud exams (GCP-ACE, GCP-CDL, GCP-PCA, GCP-PSE, GCP-PDE, GCP-GAIL)

9 Question Patterns People Get Wrong Across the Google Cloud Exams, and the Rule Behind Each

1. "Add a replica" to recover from a mistake

Where: GCP-PCA, GCP-PDE, GCP-CDL. The trap: someone dropped a table or ran a bad update, and an option offers a cross-region replica, a high-availability standby or a faster failover. The rule: replicas copy every change, including the wrong one, usually soon after it happens. They protect against losing a machine, a zone or a region, not against bad data. Recovery from a mistake needs an earlier state: point-in-time recovery, backups (immutable ones against ransomware), BigQuery time travel or snapshots, Cloud Storage versioning. Our Cloud Digital Leader explanations put it in one line: redundancy keeps the service up, replication and backups keep the data safe.

2. Expecting a new guardrail to fix what already exists

Where: GCP-ACE, GCP-PSE. The trap: an organisation policy (allowed locations, no service account keys, a custom constraint) was set, and the question asks why old resources still break it. The rule: organisation policies are generally not retroactive. They stop new violations; existing resources keep running and must be found (Cloud Asset Inventory, Security Command Center findings) and fixed. The complete design pairs a preventive control with a detective one.

3. Removing an inherited permission lower in the hierarchy

Where: GCP-ACE, GCP-PSE, GCP-CDL. The trap: a user has a role through a folder grant, and an option removes the role from the project. The rule: allow policies only add up; effective access is the union of the resource's policy and all its ancestors'. Either change the grant where it was made, or use a deny policy, which is checked before allow policies and inherited downward. The pattern is "grant high, deny the exception".

Going deeper: three kinds of policy, three questions

Patterns 2, 3 and 7 come from mixing up policies that answer different questions. An IAM allow policy answers "may this principal call this API on this resource?" and only ever grants. A deny policy answers "is this principal forbidden from this permission, whatever it was granted?" and is evaluated first. An organisation policy answers "is this configuration allowed for anyone at all?" and does not look at who is asking. A VPC Service Controls perimeter answers "may this request move data across this boundary?" and does not care whether IAM allowed the call. When an option offers a control from the wrong family, it can be perfectly configured and still change nothing. Before reading the options, name the question the stem is really asking, and only then look for the control that answers it.

Which kind of policy answers the stem's question?
May this principal call this API on this resource?
YES →
IAM allow policy (it only ever grants)
↓ no
Is this principal forbidden this permission, whatever it was granted?
YES →
Deny policy (evaluated first, inherited downward)
↓ no
Is this configuration allowed for anyone at all?
YES →
Organisation policy (stops new violations only)
↓ no
May this request move data across this boundary?
YES →
VPC Service Controls perimeter
↓ no
OTHERWISE →
Name the question first: a control from the wrong family changes nothing
Patterns 2, 3 and 7 come from picking a control from the wrong family: match the stem's question to the row that answers it.

4. Mixing up "attach" and "impersonate"

Where: GCP-ACE, GCP-PSE. The trap: a developer needs to deploy a VM that runs as a service account, and the option grants Service Account Token Creator, or the reverse. The rule: attaching an account to a resource needs actAs, which is in Service Account User. Acting as the account yourself, with short-lived tokens, needs Service Account Token Creator. Grant either on the single service account, never on the whole project, where it would cover every account including more powerful ones.

5. Searching logs that were never written

Where: GCP-ACE, GCP-PSE, GCP-PCA, GCP-PDE. The trap: an auditor asks who read the files in a bucket last week, or a detector stays silent, and an option queries the audit logs or grants Logs Viewer. The rule: Admin Activity audit logs are always on, but Data Access audit logs are off by default for all services except some BigQuery services. If nobody enabled them, the entries do not exist, and turning them on helps only from now on. Reading them also needs Private Logs Viewer; Logs Viewer does not include them.

Beginner tip: open the IAM and Admin page of a test project, go to Audit Logs, and look at the list of services. You will see that Data Access logging is unticked for almost all of them. Then upload a file to a bucket, download it, and search Logs Explorer for the download: nothing. Enable DATA_READ for Cloud Storage, download the file again, and search again. Patterns 5 and 2 both stop being abstract after this.

6. Treating a budget as a limit

Where: GCP-ACE, GCP-CDL, GCP-PCA. The trap: a team must not spend more than an amount, and the option creates a budget with alerts. The rule: an ordinary (alerts-only) budget measures and warns; it does not stop spending unless someone, or some code reading the budget's Pub/Sub messages, acts. The exception is a spend cap budget: for a single project and one eligible service (such as Cloud Run), it pauses new usage of that service once estimated costs pass 100% of the amount. A quota limits consumption in advance. When the stem says "prevent", look for a quota, a spend cap, an automated action or a narrower permission, not an alert.

7. Using IAM to stop data leaving

Where: GCP-PSE, GCP-PCA, GCP-PDE. The trap: the worry is exfiltration by someone who has valid access (a stolen key, an insider, compromised code), and the option tightens IAM roles or adds a firewall rule. The rule: IAM decides who may call an API, not where the data may go, and a firewall cannot tell one bucket from another behind the same Google API endpoint. A VPC Service Controls perimeter blocks data moving to projects outside it even for authorised callers. It works on top of IAM, never instead of it, and changes are tested in dry run first.

8. Fine-tuning a model to teach it facts

Where: GCP-GAIL, GCP-PCA, GCP-CDL. The trap: a gen AI assistant gives out-of-date or invented answers about company policy, and the option fine-tunes the model on the policy documents. The rule: missing or changing facts point to grounding or retrieval-augmented generation, which supplies current text at question time and can cite sources. A stable style or format points to fine-tuning. Prompt engineering comes before both, because it is the cheapest and needs no training data.

9. Choosing the biggest tool to be safe

Where: GCP-PDE, GCP-PCA, GCP-CDL. The trap: the requirement is modest (one nightly SQL statement, a regional HR database, a quick summary feature) and the option offers the most powerful product: Cloud Composer (now Managed Service for Apache Airflow), Spanner, a custom-trained model. The rule: the exams reward the simplest service that meets every stated requirement. One SQL statement on a schedule is a scheduled query; a regional, moderate-scale MySQL or PostgreSQL app is Cloud SQL with high availability; a common task with a deadline and no ML staff is a ready-made API or assistant. Over-buying is wrong for the same reason under-buying is: it does not match the requirement.

How to use this list

When you get a practice question wrong, check whether it belongs to one of these patterns before you look at the product details. If it does, write the rule in your own words next to the question. The pattern will return in the next set with different products, and the rule, not the product, is what you need to recognise.

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.

Rules as described in Google Cloud's documentation and the official exam guides for each exam. Defaults such as which audit logs are enabled can change; check the current documentation.