Google Associate Cloud Engineer Practice Test 24
Question 230
A non-Python agent has already been built and tested as an OCI container, stored in Artifact Registry, and implements Agent Runtime's server contract. Which deployment method retains that image?
- Use only the Python-local-object SDK execution path, which accepts arbitrary non-Python images without a container contract.
- Upload the image digest into Secret Manager and rely on Secret Manager to launch the agent service.
- Deploy the Artifact Registry container image through Agent Runtime's supported container-image configuration.
- Use Developer Connect source deployment, which is the documented non-Python entrypoint for any compiled language.
Explanation
Correct Answer
C:Deploy the Artifact Registry container image through Agent Runtime's supported container-image configuration. Agent Runtime supports Artifact Registry container images built with any language when the image satisfies its runtime contract.
Incorrect Answers
A:Use only the Python-local-object SDK execution path, which accepts arbitrary non-Python images without a container contract. The in-memory object SDK path targets supported Python agent objects and is not a generic hosting mechanism for arbitrary OCI images.
B:Upload the image digest into Secret Manager and rely on Secret Manager to launch the agent service. Secret Manager holds sensitive configuration and does not execute containers or provide a managed agent service endpoint.
D:Use Developer Connect source deployment, which is the documented non-Python entrypoint for any compiled language. The Developer Connect source path is Python-only in the documented deployment matrix; the already-built non-Python container fits the image method.
Question 231
An agent needs custom operating-system packages and an explicitly controlled API-server entrypoint. The team wants Agent Runtime to build the image from its authored Dockerfile. Which route fits?
- Deploy a Python execution object and assume the platform installs arbitrary OS libraries missing from its default build.
- Use Agent Runtime's Dockerfile/source-packages method with the required libraries and a container that satisfies the runtime contract.
- Choose a Cloud Monitoring uptime check as the Dockerfile builder because it controls the API-server entrypoint.
- Change the agent's display name; a different name automatically installs custom system dependencies.
Explanation
Correct Answer
B:Use Agent Runtime's Dockerfile/source-packages method with the required libraries and a container that satisfies the runtime contract. The Dockerfile deployment method permits control over the container image and API server while requiring the documented Agent Runtime contract.
Incorrect Answers
A:Deploy a Python execution object and assume the platform installs arbitrary OS libraries missing from its default build. An in-memory agent object does not specify the custom container build steps required to install the requested operating-system packages.
C:Choose a Cloud Monitoring uptime check as the Dockerfile builder because it controls the API-server entrypoint. An uptime check probes a running endpoint, not an image build environment or application startup command.
D:Change the agent's display name; a different name automatically installs custom system dependencies. Display names are metadata and cannot change the operating-system libraries or server process inside a deployed agent container.
Question 232
A developer prototyped an in-memory agent object in a notebook and wants a supported direct deployment route for testing. What mechanism fits?
- Package the notebook’s agent as a VM boot disk and create a MIG before evaluating any supported agent service.
- Paste the agent code into Cloud Billing budget comments so Google Cloud runs it on demand.
- Create a Kubernetes StatefulSet just to serialize an in-memory notebook agent object.
- Use Agent Platform SDK deployment of the supported in-memory local_agent object to Agent Runtime.
Explanation
Correct Answer
D:Agent Platform SDK can deploy supported in-memory agent objects, including suitable interactive notebook prototypes, to the managed Agent Runtime. Validate object serializability, dependencies and project permissions before calling the SDK deployment API.
Incorrect Answers
A:A VM image workflow may host custom code but is not the intended direct SDK deployment path from an in-memory agent object.
B:Cloud Billing budget annotations are financial metadata; they cannot run agent logic or create an endpoint.
C:Kubernetes infrastructure is not required merely to deploy a compatible in-memory agent through the managed SDK.
Question 233
A remote agent deployment starts but fails when calling a protected Google Cloud API because its runtime identity lacks read permission. Which two-stage check is needed first?
- Grant the notebook user full Organization Owner regardless of the agent’s deployed execution identity.
- Identify the Agent Runtime execution service account and grant narrowly scoped API read permissions at the target resource.
- Increase the agent’s maximum CPU allocation because IAM permission denied is always resource exhaustion.
- Disable the protected Google API’s authentication requirement for all callers.
Explanation
Correct Answer
B:A running agent still needs authorization for each downstream API. Inspect which service identity executes its tools, check the denied permission and resource, and grant only the required access without modifying the agent’s public invocation policy.
Incorrect Answers
A:The user who deploys an agent and the runtime identity used by its calls may be different principals; broad owner access is inappropriate.
C:Additional agent runtime compute capacity does not satisfy a denied permission to a protected API.
D:Disabling authentication would expose the API rather than assigning the necessary permission to the intended runtime principal.
Question 234
An agent is deployed successfully but cannot be invoked because its authorized caller lacks endpoint access. What must be configured?
- Rebuild the container image repeatedly; an authorization failure means the packaged code is invalid.
- Grant the intended caller permission to invoke the deployed agent resource, then test an authenticated query.
- Expose the agent’s endpoint to anonymous users as a permanent workaround.
- Grant the caller Cloud Billing Account Viewer because that role implicitly invokes all agents.
Explanation
Correct Answer
B:Check the deployed agent’s resource identity, caller credentials and operation-level permissions, assign the least-privilege invocation access and confirm the intended request succeeds while an unauthorized one is rejected.
Incorrect Answers
A:A successfully deployed runtime can still deny a separate caller; rebuilding does not supply missing endpoint IAM.
C:Unauthenticated access weakens security and is unnecessary for a known authorized caller.
D:Billing visibility is unrelated to the Agent Runtime invocation permission.
Question 235
A CI pipeline deploys an agent from local source on every release and the team wants immutable traceability. What release artifact should be captured?
- Keep only the latest notebook cell output, because deployed agents are always identical to the current notebook.
- Record immutable source commit, pinned dependency manifest and deployed Agent Runtime resource identifier/version in the release receipt.
- Record only the project display name, which uniquely identifies every historical agent deployment.
- Skip dependency pinning and assume managed hosting always uses the same library versions.
Explanation
Correct Answer
B:Store a source revision, relevant packaging or image digest, dependency versions and the resulting remote resource/version in a release record. This makes the deployment reproducible and supports rollback, debugging and audit comparisons.
Incorrect Answers
A:Notebook output does not identify the exact source and environment used for an earlier deployment.
C:A project name can contain many separate agent deployments and does not identify the release artifact.
D:Unpinned library changes can alter behavior between builds despite identical application logic.
Question 236
A customer agent has low idle traffic, but the team proposes manually managing an always-on CPU cluster solely for hosting the agent. What service alternative is relevant?
- Evaluate managed Agent Runtime hosting and scaling for the supported agent instead of operating a permanent self-managed cluster by default.
- Run a permanent single-node Kubernetes cluster and disable monitoring to guarantee lower total costs.
- Create a new VM for every agent request and manage cleanup manually.
- Use BigQuery SQL scheduled queries as an interactive agent RPC server with no hosted runtime.
Explanation
Correct Answer
A:Agent Runtime abstracts agent serving infrastructure and can handle managed scaling. Evaluate its framework compatibility, latency and billing profile against low traffic before committing to a manually maintained Kubernetes fleet.
Incorrect Answers
B:A self-managed always-on cluster adds operations and its relative cost cannot be presumed lowest for sparse usage.
C:Manual per-request VM creation introduces provisioning latency and additional orchestration instead of a managed agent endpoint.
D:Scheduled queries do not provide an interactive agent service with tool execution and conversational request handling.
Question 237
A team needs to test a deployed LangGraph agent remotely after publishing it to Agent Runtime. What should it verify before client rollout?
- Verify only that the agent resource name exists; its tool calls will be validated automatically on first customer request.
- Skip remote testing if a local notebook invocation succeeded before deploying dependencies.
- Grant all users invocation access so any possible framework mismatch will be visible.
- Invoke the deployed LangGraph agent with authorized credentials and representative tool-dependent requests, checking response and error paths.
Explanation
Correct Answer
D:Use the documented Agent Platform client to query the remote agent with inputs reflecting real LangGraph workflows. Verify permissions, tool calls, state handling where applicable, latency and error behavior before directing clients to it.
Incorrect Answers
A:A resource ID proves deployment occurred, not that runtime dependencies, tools or response schemas work.
B:Local success does not establish that the remote deployment has matching dependencies, identities and network access.
C:Public invocation is not a functional testing requirement and expands the attack surface.
Question 238
A GKE Autopilot Pod is rejected because its memory request exceeds the selected compute class maximum. What should the operator change?
- Ignore the admission error and rely on the scheduler to provide memory larger than the compute class supports.
- Reduce requests to supported bounds or choose a compute class/hardware configuration with sufficient allowed memory.
- Set the Pod’s memory limit to unlimited and leave an invalid memory request unchanged.
- Add a Kubernetes Service to the Pod; Service objects override Autopilot resource limits.
Explanation
Correct Answer
B:Select a class whose documented CPU, memory and ratio limits cover the workload, or right-size the requested memory. An over-maximum request requires a valid manifest or compatible class rather than repeated scheduling attempts.
Incorrect Answers
A:Autopilot validates requested resources and can reject requests above supported maximums.
C:An unspecified limit does not make an out-of-range resource request valid for the chosen class.
D:Services provide networking; they do not change admission-controller resource validation.
Question 239
A Pod on GKE Autopilot starts with unexpected default CPU requests when no requests were declared. What explains the behavior?
- Assume the Pod is using exactly zero CPU because the manifest omitted requests.
- Review Autopilot’s default request assignment and explicitly set supported CPU/memory requests for the workload.
- Disable the Kubernetes scheduler because it assigns CPU only after a Pod is running.
- Move the Pod to a different namespace; all unspecified resource requests disappear there.
Explanation
Correct Answer
B:Inspect the admitted Pod spec and applicable compute-class defaults. Declare measured, supported resource requests to make scheduling and billing more predictable instead of depending on automatically applied defaults.
Incorrect Answers
A:Autopilot can set default CPU requests when the manifest omits them, so no field does not mean no reserved resources.
C:Resource defaults can be applied at admission; disabling scheduling cannot make the Pod usable.
D:Moving the Pod to another namespace does not automatically bypass GKE Autopilot resource request defaults.
Question 240
A GKE Autopilot service uses an Arm-only container image in a supported region. Its Pod has no architecture selection and starts on x86 nodes. What should the team specify?
- Set a VPC route priority of arm64; node architecture follows the network route chosen by the Pod.
- Specify a compatible Arm compute class or supported kubernetes.io/arch: arm64 scheduling selector for the cluster version.
- Increase the CPU limit without adding node architecture attributes; Kubernetes converts x86 hardware to Arm automatically.
- Set the Cloud Storage object metadata architecture=arm64; the scheduler uses bucket metadata to select nodes.
Explanation
Correct Answer
B:Specify a compatible Arm compute class or supported kubernetes.io/arch: arm64 scheduling selector for the cluster version. The documented Autopilot Arm deployment paths use architecture selectors or an eligible Arm compute class. Matching the image to an Arm node avoids startup incompatibility.
Incorrect Answers
A:Set a VPC route priority of arm64; node architecture follows the network route chosen by the Pod. Routing determines traffic paths, not instruction-set architecture or kube-scheduler placement.
C:Increase the CPU limit without adding node architecture attributes; Kubernetes converts x86 hardware to Arm automatically. CPU quantity does not specify architecture. An Arm-only image cannot run on an incompatible x86-only node merely by requesting more cores.
D:Set the Cloud Storage object metadata architecture=arm64; the scheduler uses bucket metadata to select nodes. Cloud Storage object labels are not Kubernetes node scheduling constraints and cannot direct the Pod to an Arm node.