cert-bank

Guides

이 글을 한국어로 읽기

GCP-ACE Study Guide: Build One Project From Empty to Production, and Follow IAM All the Way

The Associate Cloud Engineer exam is one of the most hands-on of the Google Cloud certifications. It asks what you would type, click or grant next. The easiest way to study for it is to live through the life of one project: create it, pay for it, put workloads on it, keep them running, and decide who may touch them. This guide follows that order, and it points out where permissions come back at every stage, because in this exam they always do.

Study guide · Google Cloud Certified Associate Cloud Engineer (GCP-ACE)

GCP-ACE Study Guide: Build One Project From Empty to Production, and Follow IAM All the Way

The exam

50 to 60 multiple choice and multiple select questions in 2 hours. Google does not publish a pass mark: you get a pass or a fail, so check the official exam page for the current details and aim well above any guessed threshold. Four sections, with approximate weights: Setting up a cloud solution environment (20%), Planning and implementing a cloud solution (30%), Ensuring the successful operation of a cloud solution (30%) and Configuring access and security (20%).

Two figures from the 275 GCP-ACE questions in this question bank explain the angle of this guide. IAM is the most frequent service by far: it appears in 76 questions, and 21 of those sit outside the access and security section. And 78 questions (28%) mention gcloud, kubectl or bq in the stem or the options, nearly all of them as a full command. You will not have to write commands from memory, but you must be able to read one and see what it does.

Stage 1: an empty organisation and a bill

Everything starts with the resource hierarchy: organisation, folders, projects, resources. Two kinds of rules flow down it, and the exam keeps them apart:

  • IAM allow policies decide who can act. Grants add up down the tree: a role on a folder reaches every project below it.
  • Organization Policy decides what configurations are allowed for everyone, project owners included: allowed locations, no default network, no service account keys.

A new organisation policy usually does not touch resources that already exist. If a question says "a constraint was set yesterday, why is last month's VM still in the wrong region?", that is the answer, and the fix is to find the old resources (for example with Cloud Asset Inventory) and correct them by hand.

Billing is the other half of day zero. Know which billing role does what: Billing Account User only lets someone link projects to the account, Billing Account Costs Manager handles budgets and the cost export without controlling links, Billing Account Viewer reads costs for finance, and Project Billing Manager links or unlinks one project. The trap everyone meets once: an alerts-only budget sends emails but never stops spending. To act automatically you connect the budget to a Pub/Sub topic and let code respond.

Beginner tip: create a free-trial project and do stage 1 yourself before reading further. Create a budget with a small amount and the default 50/90/100% thresholds, then open IAM and give a second Google account the Viewer role on the project. Sign in as that account and try to create a Cloud Storage bucket. The error message you get, naming the missing permission, is the situation many troubleshooting questions describe in words, and seeing it once makes them easy to picture.

Stage 2: putting workloads on it

The planning section is a series of "which product fits" choices, each with one deciding phrase:

The stem saysThink of
Specific OS, drivers, a legacy app moved as isCompute Engine; a managed instance group from an instance template for scaling and autohealing
Stateless container, scale to zero, little to operateCloud Run
Many containers, Kubernetes skills, node controlGKE Standard; GKE Autopilot when Google should manage the nodes
Run code when a file lands or a message arrivesCloud Run functions, with Eventarc or Pub/Sub
Batch work that can be interruptedSpot VMs

Two naming changes cost points if you learned from old material: Container Registry has shut down (the answer is Artifact Registry), and Cloud Deployment Manager has reached end of support (the answer is Infrastructure Manager or Terraform).

Permissions return here too. A Cloud Run service requires every caller to hold the Cloud Run Invoker role unless you disable the invoker check; for a public website, disabling that check is the documented route, because organisations with domain restricted sharing reject a grant to allUsers. A VM runs as the service account you attach, and attaching one needs a role on that service account.

Going deeper: attach versus impersonate, and why the role sits on the service account

A service account is both an identity and a resource with its own allow policy. Attaching it to a VM or a Cloud Run service needs the iam.serviceAccounts.actAs permission, which is in Service Account User. Acting as it yourself, from your laptop or a script, means asking for short-lived tokens, which needs Service Account Token Creator. Grant either role on the single account, not on the project: a project-level grant covers every service account in the project, including more powerful ones, which turns a small grant into a path to escalate privileges. That is why the right answer so often reads "grant the role on the service account" rather than "on the project". Short-lived tokens from impersonation are also why the exam prefers it to downloading a key: a key works from anywhere and does not expire unless a policy makes it.

Stage 3: keeping it running

This section shares the largest weight (30%) with planning, and it is the one with the most troubleshooting questions in our bank (29 of 83). Learn the operational tools by the question they answer:

  • "Who changed this?" Admin Activity audit logs, always on. "Who read this data?" Data Access audit logs, which are off by default for every service except some BigQuery services. If nobody enabled them, the entries never existed; enabling them now helps only from now on.
  • "Tell me when it breaks." An alerting policy with a metric-threshold condition. "Tell me when it goes silent." A metric-absence condition, which fires when data stops arriving.
  • "Stop storing this noise." An exclusion filter on the _Default sink. "Send logs elsewhere." A sink to BigQuery, Cloud Storage, Pub/Sub or another project; the sink's writer identity needs a role on the destination.
  • "Someone deleted the object." Soft delete (on by default for new buckets, 7 days) restores it; Object Versioning keeps older versions only from the moment it was enabled.
  • "Traffic is blocked." Firewall rules: the lowest priority number wins, deny beats allow only at equal priority, and routes choose a path but never allow traffic.

Expect commands here. A question may show gcloud compute firewall-rules update RULE --disabled and ask what happens: the rule keeps its settings but is not enforced.

Stage 4: deciding who may do what

The access section pulls the thread together. The rules worth writing on a card:

  1. Least privilege is the narrowest role on the smallest resource: Pub/Sub Publisher on one topic, not Editor on the project.
  2. Effective access is the union of the resource's policy and every ancestor's. Removing a project grant changes nothing while the same role is granted on the folder.
  3. Owner, Editor and Viewer are now legacy basic roles and cannot take IAM Conditions; the current basic roles are Admin, Writer and Reader. Prefer predefined roles in production.
  4. Custom roles exist at organisation or project level, never on a folder.
  5. Deny policies are checked before allow policies, and no allow role, not even Owner, overrides them.
Going deeper: why you cannot remove an inherited permission lower down

IAM allow policies only ever add. When Google Cloud checks a request, it collects the allow policies of the resource and all its ancestors and asks whether any of them grants the permission. A lower level has no way to say "except here", so a folder-wide grant cannot be cancelled on one project. There are two real fixes: change the grant where it was made, or attach a deny policy, which is evaluated first and inherited downward. The documented pattern is "grant high, deny the exception". Deny conditions recognise only resource tags, so to exempt production projects you tag them and write the condition on the tag. Managing deny policies needs its own role (Deny Admin), which Owner does not include.

One resource hierarchy, two kinds of rules flowing down it
Organisation
Organization Policy: which configurations are allowed · Usually does not touch existing resources · Custom roles can live here
Folders
A role granted here reaches every project below · No custom roles
Projects
Linked to a billing account · Custom roles can live here · Removing a grant does nothing if the folder grants it
Resources
Effective access: union of its own and every ancestor's allow policy · A deny policy beats any allow, even Owner
Allow grants and policies flow from top to bottom; a lower level cannot remove what a higher level grants.

A study order

  1. Hierarchy, organisation policies, billing roles and budgets (stage 1), in a real project.
  2. IAM and service accounts, early, because every later topic leans on them.
  3. Compute, GKE, Cloud Run, storage classes and databases.
  4. VPC networks, firewall rules, Cloud NAT, Cloud DNS and load balancers.
  5. Logging, monitoring and backups, then timed mixed practice sets.

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.

Section weights as listed in Google's Associate Cloud Engineer exam guide. Question count, product names and exam rules change; check the current official exam page before you book.