Free Cloudera Generative AI Engineer Sample Questions

Correct answer: A — Each tree is trained on a bootstrap sample of the rows and a random subset of features at each split, and predictions are averaged or voted across trees to cancel out individual trees' errors

• A. Bootstrap aggregating (bagging) combined with random feature subsampling at each split decorrelates the individual trees, and averaging or majority-voting their predictions reduces the variance that makes a single tree unstable across resamples.
• B. Random forests still use recursive axis-aligned splits in every tree; there is no global linear decision boundary replacing the tree structure, so this does not describe how the ensemble achieves stability.
• C. Training every tree on the identical dataset and keeping only the best performer would not reduce variance at all, since it discards the diversity that bagging depends on and risks selecting the tree that most overfits the training data.
• D. Random forests do not require a PCA preprocessing step; the randomness that stabilizes the ensemble comes from bootstrap row sampling and random feature selection at each split, not from projecting onto principal components.

Correct answer: A — K-means starts from randomly placed initial centroids and can converge to different local optima, so running the algorithm multiple times with different initializations and keeping the result with the lowest within-cluster variance reduces this variability

• A. K-means begins with randomly chosen (or randomly seeded) initial centroids, and different starting positions can lead the algorithm to converge to different local optima; running it several times with different initializations and selecting the lowest within-cluster sum of squares is the standard mitigation, often implemented as multiple restarts.
• B. K-means does not infer `k` from the data on its own; the number of clusters is a fixed hyperparameter that the user must supply before running the algorithm, so this does not explain the observed variability.
• C. K-means uses a fixed distance measure, typically Euclidean distance, throughout a run rather than generating a new metric randomly each time, so this is not the source of the run-to-run differences.
• D. K-means does not randomly reorder feature columns between runs; the variability in cluster assignments comes from the random initialization of centroids, not from any shuffling of the input features.

Correct answer: A — Recall increases because more actual fraud cases are captured, while precision decreases because more legitimate transactions are also flagged

• A. Lowering the threshold means transactions with weaker fraud signals now get flagged, which pulls in more of the true fraud cases and raises recall. It also sweeps in more legitimate transactions as false positives, which pushes precision down.
• B. A lower threshold flags more transactions overall, which tends to catch more true fraud cases and raise recall, not lower it, while precision typically falls rather than rises as weaker signals get included. This describes the opposite of the actual relationship.
• C. Lowering the threshold does not raise the model's confidence in its predictions; it simply changes which predictions get labeled positive. Precision and recall move in opposite directions in this scenario rather than both improving.
• D. The threshold change affects which predictions are labeled positive, not how many transactions the model evaluates in total. Both metrics do not decrease together under this kind of threshold adjustment.

Correct answer: A — Use an iterative imputer that models each feature with missing values as a function of the other features and repeats the estimates until they converge

• A. An iterative imputer repeatedly regresses each feature with missing values on the other available features, which captures the correlation between debt_ratio, income, and credit_utilization rather than treating each column in isolation.
• B. Filling every missing entry with a constant zero ignores the relationships between the correlated features entirely and can introduce an artificial spike in the data that does not reflect any real correlation structure.
• C. Most-frequent-value imputation applied column by column is a univariate strategy that, like mean imputation, ignores cross-feature relationships and does not model why income and debt_ratio tend to be missing together.
• D. Dropping every record with any missing value among the three correlated features discards data rather than modeling the interdependency, and can bias the remaining dataset if the missingness itself is systematic.

Correct answers: A, B, C — StandardScaler computing the per-feature mean and standard deviation; SimpleImputer computing the mean value used to fill missing entries; TargetEncoder computing per-category target means for a categorical feature

• A. StandardScaler learns the mean and standard deviation directly from the data it is fit on, so those statistics must come only from the training split to avoid incorporating test-set values.
• B. SimpleImputer's mean-fill value is computed from the data it is fit on, so it must be learned only from the training split; fitting it on the full dataset would let test-set values influence the imputed training data.
• C. TargetEncoder derives per-category statistics from the target values in the fitted data, so those category means must be learned only from the training split to avoid leaking target information from the test set.
• D. Binarizer applies a threshold chosen in advance by the analyst rather than a statistic learned from the data, so there is no data-dependent fitting step that could leak test-set information.
• E. Stripping whitespace with a regex is a fixed, stateless string operation that does not compute any statistic from the dataset, so it carries no risk of leaking test-set information.
• F. Mapping categories through a fixed, pre-defined dictionary does not derive its values from the dataset being processed, so there are no learned statistics that could leak information between splits.

Correct answer: A — A Session provides an interactive, ad hoc compute environment, while a Job runs code on a defined or recurring schedule

• A. Sessions give data scientists an interactive compute environment for exploratory work, while Jobs execute project code on a schedule or trigger, correctly matching each concept to its role.
• B. This swaps the two definitions: interactive, ad hoc work happens in a Session, and scheduled or recurring execution is what a Job performs, not the reverse.
• C. A Project is the top-level container holding code, files, and configuration, and a Session is one interactive compute environment launched within a project, so the hierarchy here is inverted.
• D. An Application is the long-running, browser-accessible web service in CAI, while a Job is the scheduled or batch task, which is the opposite of how this option pairs them.

Correct answers: A, B, C, D — Compute resources, including CPU/GPU instance types and autoscaling ranges; Networking, including subnet selection and load balancer settings; Governance, including enabling Apache Atlas integration and model metrics tracking; Security, including optional TLS and public internet access restrictions

• A. Compute resource configuration, covering CPU and GPU instance types along with autoscaling ranges and root volume sizes, is a documented advanced provisioning category.
• B. Networking settings, including subnet configuration and load balancer options, are part of the advanced configuration available during workbench provisioning.
• C. Governance settings, such as enabling Apache Atlas integration and model metrics tracking, are exposed as an advanced toggle during provisioning.
• D. Security options, including optional TLS termination and public internet access or IP range restrictions, are part of the advanced provisioning configuration.
• E. Workbench provisioning does not include a billing category for selecting payment methods or spend caps; billing is handled outside the workbench creation workflow.
• F. There is no advanced provisioning category for enforcing a Git branching strategy across projects; source control is managed within individual projects, not at workbench provisioning time.

Correct answer: A — Editor, kernel, edition, and version

• A. Correct. The Runtime Catalog lists each available ML Runtime by its editor (such as JupyterLab), kernel, edition, and version, letting a user select the precise image needed for a project.
• B. Node pool, availability zone, and instance type describe underlying Kubernetes infrastructure and Resource Group configuration, not the attributes used to select an ML Runtime image in the catalog.
• C. Resource profile name, vCPU count, and memory allocation describe a Compute Resource Profile, which controls CPU/GPU/memory reservation for a workload, not the runtime image identity.
• D. Workspace tier, provisioning region, and storage class relate to workspace provisioning decisions made through the Management Console, not to selecting an individual runtime image.

Correct answer: A — A Docker image, either a Legacy Engine or an ML Runtime, containing the OS, interpreters, and libraries used to execute user code, instantiated as a virtual-machine-style environment for the duration of the workload.

• A. Correct. An engine refers to the Docker image, whether a Legacy Engine or an ML Runtime, that provides the OS, interpreters, and libraries and is spun up as a virtual-machine-style environment each time a session, job, experiment, model, or application runs.
• B. An engine is a software image rather than a dedicated physical GPU node; GPU allocation is handled separately through Compute Resource Profiles and cluster scheduling, not by permanently reserving hardware to one project.
• C. The internal Postgres database stores platform metadata in the data tier of the three-tier architecture; it is a distinct component from the engine, which is the runtime execution environment for user code.
• D. The editor (JupyterLab, Workbench, etc.) is one attribute exposed by a runtime in the Runtime Catalog, but the engine itself is the underlying container image, not merely the front-end editor interface.

Correct answer: A — The project directory in the container is backed by the platform's internal NFS-based persistent mount, while other paths in the container filesystem are ephemeral and local to that pod.

• A. This is correct: the project directory is served from the platform's internal NFS mount, which persists independently of any single pod, while paths outside the project directory live only in that container's local, ephemeral filesystem and disappear when the pod is rescheduled.
• B. The internal Postgres database stores platform metadata such as user, project, and job records, not project file contents, so it has no role in replicating or dropping files based on their path.
• C. Source-to-Image builds run when a runtime or engine image is constructed, not on every session start, and they do not selectively discard files written at runtime to paths outside the project directory.
• D. There is no automatic compression-and-deletion mechanism for files outside the project directory; the actual cause is simply that those paths are not backed by persistent storage at all.

Correct answer: A — Register the corporate proxy's host, port, and protocol as a shared resource in the Management Console, then reference that proxy when configuring the environment.

• A. This is correct: proxies are registered as shared resources with host, port, and protocol details in the Management Console and then referenced when creating or updating an environment, which is the documented way to control outbound connections.
• B. Manually adding a proxy address to per-project environment variables is not the documented mechanism and would require every process to individually honor the setting rather than having it enforced platform-wide.
• C. Installing a SOCKS5 client inside every session image is not how outbound proxying is configured on this platform, and it would bypass the centralized proxy registration and control the security team needs.
• D. An ingress controller filters inbound traffic to services, not outbound internet connections initiated from within the workbench, so it does not address this requirement.

Correct answer: A — Cloudera AI integrates a native plugin acting as an interface between its API and the MLflow SDK, using the MLflow client library as the default logging method.

• A. Cloudera AI ships a native plugin that bridges its own API with the MLflow SDK, and it uses the standard MLflow client library as the default way to log parameters, metrics, and artifacts for experiments.
• B. Cloudera AI is compatible with the MLflow tracking API rather than replacing it with an incompatible custom syntax, which is precisely what lets standard MLflow client calls work in the platform.
• C. MLflow tracking support in Cloudera AI covers both traditional ML experiment logging and LLM evaluation workflows through the same `mlflow.evaluate()` and tracking API, not just one narrow use case.
• D. Tracking works against the workspace's built-in tracking store without requiring users to stand up and manage a separate external MLflow tracking server before logging is possible.

Correct answers: A, B, C — A model to evaluate, supplied as an MLflow pyfunc model, a registered model URI, or a Python callable such as a Hugging Face pipeline.; A set of metrics to compute, which may be heuristic-based, LLM-as-a-Judge based, or a combination of both types.; Evaluation data, such as a pandas DataFrame, Python list, numpy array, or MLflow Dataset instance containing inputs and references.

• A. A model reference is one of the three core inputs `mlflow.evaluate()` needs, and it explicitly accepts pyfunc models, registered model URIs, or plain Python callables like a Hugging Face pipeline.
• B. Metrics are a required input to the evaluation call, and the documentation describes both heuristic-based metrics and LLM-as-a-Judge metrics as valid, combinable choices for scoring.
• C. Evaluation data is the third required component, and it can be supplied in several accepted formats including pandas DataFrames, lists, numpy arrays, or MLflow Dataset objects.
• D. A Registry deployment endpoint is not one of the required inputs to `mlflow.evaluate()`; the function can evaluate a model reference directly without any prior deployment step through the Registry.
• E. No dedicated GPU-backed inference service is required as a prerequisite for calling `mlflow.evaluate()`; the evaluation runs against the supplied model reference using the three core inputs already described.

Correct answer: D — Convert the trained scikit-learn model into ONNX format, then register the resulting ONNX artifact as a new model version in the Registry

• A. The Registry stores whatever artifact format is registered; it does not perform automatic framework-to-ONNX conversion on deployment, so a scikit-learn artifact registered as-is would remain scikit-learn.
• B. A hyperparameter file describes training configuration, not a servable model artifact, so registering it in place of the trained model would leave no usable inference artifact for the ONNX runtime.
• C. Fine Tuning Studio is built around adapting large language models with PEFT techniques, not converting arbitrary scikit-learn estimators into ONNX, so it is not the right tool for this conversion.
• D. Converting the scikit-learn model into ONNX format before registration produces an artifact compatible with the ONNX-standardized serving infrastructure, and registering that artifact creates a new tracked version.

Correct answer: A — Pre-built, expert-authored reference projects that teams can install and adapt as templates for structuring their own pipelines and deployments.

• A. Applied ML Prototypes are complete, expert-built reference solutions that teams install and adapt, giving them a working structural template for pipelines and deployments rather than building from scratch.
• B. AMPs are installable reference projects, not an orchestration service that automatically builds and deploys every project in a workspace on its own.
• C. Monitoring dashboards for Prometheus and Grafana metrics are a separate workspace capability; AMPs are reference project templates, not a metrics visualization layer.
• D. Access control over Jobs and Registry actions is governed separately from AMPs; AMPs are reference code and project templates, not a permissions mechanism.

Correct answer: A — Schedule a monitoring Job that computes drift metrics and, when the threshold is exceeded, calls the Jobs API to start the retraining Job automatically.

• A. A scheduled monitoring Job that computes drift metrics and programmatically calls the Jobs API to start retraining once a threshold is crossed is exactly how conditional, unattended retraining is built on the platform.
• B. The Inference Service serves predictions; it does not itself detect drift and trigger retraining, so waiting on it would never actually start a new training run.
• C. Changing a Compute Resource Profile only adjusts CPU or memory allocation for a workload; it has no mechanism for detecting drift or starting a retraining pipeline as a side effect.
• D. Scaling replicas to zero stops serving traffic but does not start a retraining process; there is no built-in link between replica count and triggering a new training run.

Correct answer: A — SentencePiece, which treats raw text as a character stream, encodes spaces as an explicit symbol, and applies BPE or Unigram on top

• A. SentencePiece works directly on the raw text stream without assuming whitespace-delimited words, representing spaces themselves as a vocabulary symbol before applying an underlying BPE or Unigram model, which makes it suitable for languages without word-separating spaces.
• B. Word-level tokenization still relies on identifying word boundaries to build its vocabulary, and without whitespace cues it cannot reliably segment continuous Japanese or Chinese text into words.
• C. Running byte pair encoding with no pre-tokenization step is not the standard design; standard BPE implementations still expect a pre-tokenizer to define initial word boundaries, so this does not describe how the space-free case is actually solved.
• D. Manually enumerating every Han character into a WordPiece base vocabulary does not solve the underlying problem of missing word boundaries and does not match how WordPiece itself is designed to operate.

Correct answer: A — In the system prompt, since it defines persistent behavioral constraints that should apply to every user turn regardless of message content

• A. System prompts set persistent, high-priority instructions that apply across every turn and are generally weighted above user content, making them the right place for durable formatting and confidentiality rules.
• B. Embedding the rule inside the user turn mixes untrusted end-user text with the instruction, weakening enforcement because that channel is exactly what a user could try to argue with or override.
• C. Relying only on a single few-shot example for a hard constraint like confidentiality is unreliable, since the model may generalize the pattern imperfectly instead of treating it as a firm, non-negotiable rule.
• D. Placing the rule inside the assistant's first reply means it was never available as an instruction beforehand, so the model has nothing to follow when generating that reply or any prior turn.

Correct answers: A, B, F — High request volume on a narrow task amortizes the training cost across many self-hosted calls, avoiding per-token fees; A smaller self-hosted model can serve the narrow task with lower per-request latency than a large external API; Self-hosting keeps sensitive request data inside the organization's own infrastructure instead of a third-party service

• A. This is correct because for a narrow, high-volume task, the fixed one-time cost of training a small adapter can be spread across a large number of self-hosted, low-latency requests, avoiding ongoing per-token API fees that scale with volume.
• B. This is correct because a smaller model dedicated to one narrow task typically has less computation per request than a large general-purpose model, which can translate into lower latency when self-hosted.
• C. This is incorrect because accuracy depends on task complexity, data quality, and model capability; a fine-tuned smaller model is not universally more accurate than a larger foundation model across all tasks.
• D. This is incorrect because foundation model APIs are commonly used for classification tasks via prompting; there is no inherent restriction preventing their use for this purpose.
• E. This is incorrect because a trained adapter still needs to be hosted and served through inference infrastructure such as Cloudera AI Inference; training does not remove the need to deploy and run the model.
• F. This is correct because routing every request to an external API means sensitive data leaves the organization's boundary, whereas a self-hosted adapted model keeps that data on internal infrastructure, which is a legitimate cost-and-risk factor in a total cost of ownership comparison.

Correct answer: A — The narrow JSON training data biased the adapter's added update toward that pattern even on unrelated prompts

• A. This is correct because even though the base weights stay frozen, the additive adapter update is applied on every forward pass regardless of prompt type, so an adapter heavily biased toward a narrow output format can skew generations on unrelated tasks unless the prompt clearly signals a different intent.
• B. This is incorrect because LoRA does not delete or physically alter any base model weights; the base matrices remain fully intact and frozen throughout adapter training and inference.
• C. This is incorrect because positional encoding mechanisms are part of the frozen base architecture and are not modified by LoRA adapter training, which only introduces additive low-rank updates to targeted weight matrices.
• D. This is incorrect because attaching a LoRA adapter does not change the context window size of the underlying model; context length is a property of the base architecture and serving configuration, not of the adapter.

Correct answer: A — Increase the chunk overlap so a portion of the preceding chunk's content carries into the next chunk, preserving boundary context

• A. Chunk overlap controls how much of the previous chunk's data is repeated in the next chunk specifically to prevent information at chunk boundaries from being lost, which directly matches the reported symptom.
• B. Changing the vector database backend affects where and how vectors are stored and queried, but it does not change how documents are split into chunks, so boundary information loss would persist.
• C. The summarization model enables summary-based retrieval as a supplementary retrieval path; it does not change how the underlying chunks are split or fix lost context at chunk boundaries.
• D. Reducing chunk size makes chunks smaller and more numerous, which can worsen boundary-cutoff issues rather than fix them, since more boundaries are introduced without overlap protection.

Correct answer: A — Cosine distance

• A. Cosine distance measures the angle between two embedding vectors and is the similarity metric RAG Studio applies when comparing a query vector against stored document vectors.
• B. Manhattan distance sums absolute coordinate differences and is not the metric RAG Studio's vector search uses for semantic similarity.
• C. Jaccard distance compares set overlap and is unsuited to dense embedding vectors, so it is not the metric applied here.
• D. Hamming distance compares differing positions in equal-length discrete strings, which does not apply to continuous embedding vectors used for semantic search.

Correct answer: A — Configure an external PostgreSQL database so metadata persists independently of any single application pod.

• A. An external PostgreSQL instance lives outside any single pod, so multiple replicas can share the same durable, consistent metadata store across restarts.
• B. Embedded H2 is a local, in-process database tied to a single pod's storage and has no built-in replication across other replicas, making it unsuitable for a multi-replica deployment.
• C. Embedded Qdrant is a vector database for embeddings, not the metadata database, so it does not hold knowledge base metadata records.
• D. File storage settings such as S3 govern where source documents are kept, not where structured metadata records are stored, so this does not address the metadata durability requirement.

Correct answer: C — Embedded Qdrant

• A. Apache Solr is not a vector store option offered by RAG Studio for knowledge base creation.
• B. ChromaDB is a supported option, but it must be explicitly configured for local or remote use; it is not what is applied automatically.
• C. Embedded Qdrant is the vector store RAG Studio uses locally by default when no external backend has been configured.
• D. Cloudera Semantic Search requires explicitly supplying a host, namespace, and authentication details, so it is not the automatic default.

Correct answers: A, B, C — Deployed workflows include built-in observability and logging for monitoring and troubleshooting.; Custom tool code executes in a sandbox using Linux namespace isolation rather than with elevated host privileges.; High-code mode via Cloudera AI Workbench lets developers build custom agents and tools beyond the low-code catalog.

• A. Agent Studio's full lifecycle management explicitly includes built-in observability and logging for deployed, production-ready workflows.
• B. Tool execution is sandboxed with Linux namespace isolation and runs without elevated privileges, which contains what invoked code can access.
• C. High-code mode integrates with Cloudera AI Workbench specifically to let developers build custom agents and tools that go beyond the low-code catalog.
• D. Agent workflows execute tool calls as part of their own orchestration and do not require first being redeployed as a separate Cloudera AI Model REST endpoint.
• E. Namespace isolation is the containment mechanism used during tool execution, not something disabled to grant broader filesystem access.
• F. Observability and logging are part of production lifecycle management for deployed workflows, not limited to sessions actively open in the low-code editor.

Correct answer: A — Store the service account's credentials securely and have the tool attach the bearer token to each outbound request

• A. Attaching a securely stored service-account bearer token to each outbound request is the correct pattern for a machine-to-machine tool call, matching how automated agent tools authenticate to protected internal APIs.
• B. Prompting a user to paste a personal password into chat exposes credentials to the LLM context and logs, and does not match the service-account authentication the API requires.
• C. Assuming an unauthenticated path exists contradicts the stated requirement that the API needs a bearer token, and skipping authentication would likely fail or violate access controls.
• D. Pre-loading expected responses into a knowledge base would return stale, fabricated approval data instead of performing the real, authenticated action the workflow requires.

Correct answer: A — Sequential processing, so tasks run in the defined order without a manager agent choosing who acts next

• A. Sequential processing executes tasks in the order they are defined, which matches a workflow where the collection, summarization, and formatting steps must always happen in the same fixed order.
• B. Hierarchical processing adds a manager agent that dynamically delegates tasks based on agent expertise, which is unnecessary and would introduce runtime variability where a fixed, predictable order is required instead.
• C. Agentic ETL describes a use-case pattern for agent-driven data pipelines, not a workflow execution model choice between fixed ordering and dynamic delegation.
• D. Switching to high-code mode addresses building custom agents or tools from scratch; it is not the mechanism Agent Studio uses to select between sequential and hierarchical task execution.

Correct answer: A — Each agent maintains interaction memory that it draws on alongside its role, goal, and tools while collaborating with other agents

• A. Agents are autonomous units that can perform tasks, make decisions, use tools, collaborate with other agents, and maintain interaction memory, which is what enables informed, stateful decisions across a workflow run.
• B. If agents discarded all context between tasks and relied only on the task description, they could not build on prior steps or collaborate meaningfully, which contradicts how memory-enabled agents behave in a stateful workflow.
• C. Memory is not limited to a manager agent; any agent in the workflow maintains interaction memory to inform its own decisions, regardless of whether the workflow uses hierarchical delegation.
• D. Replica count is an infrastructure autoscaling setting for a deployed endpoint and has no bearing on whether an individual agent retains conversational or task context during a run.

Correct answer: A — Ranger Authorization Service (RAZ)

• A. This service extends Ranger's fine-grained policy authorization and audit logging to cloud object stores like S3 and ADLS, applying the same access control model used for HDFS to cloud storage paths.
• B. This gateway acts as a reverse proxy and single access point for REST and HTTP APIs into the cluster, but it does not evaluate Ranger resource policies against cloud object store paths.
• C. This service manages encryption keys for HDFS transparent data encryption zones, which is a separate concern from fine-grained authorization over object store paths.
• D. This governance tool tracks metadata, classifications, and lineage across data assets, but it does not itself enforce access decisions against cloud object storage.

Correct answer: A — A glossary term, which provides one standardized business definition associated with multiple entities so searches under either name locate the same model.

• A. This is correct. A glossary term establishes a single standardized business definition that can be linked to multiple entities, resolving exactly the naming inconsistency described between the two business units.
• B. This is incorrect. A classification tag is oriented toward driving access control policies rather than reconciling two different business names for the same underlying model.
• C. This is incorrect. A business metadata attribute stores a custom field on a single entity instance and does not provide the shared vocabulary needed to unify two different names across teams.
• D. This is incorrect. A lineage relationship records how the model was produced from upstream data and has no role in reconciling inconsistent business terminology used to refer to the model.

Correct answer: A — Knox Gateway acts as a reverse proxy that encapsulates Kerberos authentication within the cluster, so external clients interact with REST endpoints without direct Kerberos configuration.

• A. This is correct: the gateway sits between external clients and cluster services, handling Kerberos negotiation internally so REST/HTTP clients never need to manage tickets themselves.
• B. URL rewriting is a real Knox feature but it maps request paths to backend service URLs; it does not eliminate the need for authentication on secured clusters.
• C. Knox does not issue long-lived API keys that bypass Kerberos entirely; it abstracts Kerberos handling rather than replacing it with a separate bypass credential.
• D. Authentication abstraction is a function of the gateway itself, not something delegated to the Cloudera AI Registry, which serves a different governance purpose.

Correct answer: A — UMS JWT issued by the Cloudera Control Plane

• A. The UMS JWT, issued by the Control Plane, is the token type required for Model Endpoint management operations within the Registry.
• B. The Data Lake Knox JWT is environment-scoped and better suited to inference calls and general programmatic access, not Registry endpoint management specifically.
• C. The auto-generated Kerberos JWT is tied to interactive Workbench sessions and is not the credential used for Registry endpoint management.
• D. API keys restrict access to individual deployed models via the Authorization header; they are unrelated to managing endpoints within the Registry.

Correct answer: A — NVIDIA NIM

• A. Correct. NVIDIA NIM provides pre-built, GPU-optimized containers for supported text-generation and embedding models with minimal manual tuning, and it exposes OpenAI-compatible endpoints suited to this scenario.
• B. NVIDIA Triton targets deep-learning models via ONNX backends and is not the pre-packaged, OpenAI-compatible NVIDIA container runtime the team is looking for.
• C. Hugging Face Transformers loads native Hugging Face artifacts directly and requires the team to manage runtime tuning themselves; it is not a pre-built NVIDIA-optimized container.
• D. vLLM is an open-source high-throughput LLM serving engine rather than a pre-built NVIDIA container, so it does not match the requirement for a minimal-tuning NVIDIA-optimized deployment.

Correct answers: A, B, C — Set the Endpoint Autoscale Range minimum above zero so the endpoint avoids cold-start delays during business hours; Choose concurrency per replica as the autoscaling trigger metric so scaling reacts to actual per-replica load rather than raw request counts; Select the latency profile with higher tensor parallelism to minimize Time to First Token and Inter-Token Latency

• A. Keeping a nonzero minimum during business hours ensures at least one warm replica is always ready, avoiding the multi-minute cold-start delay that comes from scaling up out of zero replicas.
• B. Concurrency per replica reacts to how much actual work each replica is handling, letting the endpoint add capacity before individual replicas become overloaded, which supports lower response times.
• C. The latency profile is specifically designed to minimize Time to First Token and Inter-Token Latency by using more GPUs per replica through higher tensor parallelism, matching the latency-sensitive requirement.
• D. A permanent minimum of zero guarantees the endpoint will hit a cold start on every idle-to-active transition, which directly undermines the goal of minimizing user-facing latency.
• E. The throughput profile trades away Time to First Token to reduce GPU count, which works against a latency-sensitive customer-facing use case rather than supporting it.
• F. A fixed single replica cannot absorb demand spikes regardless of how many worker nodes are available, since it is model endpoint autoscaling, not node autoscaling, that adjusts replica counts to match request load.

Correct answer: A — The Quota Management feature has not been enabled in the Cloudera control plane.

• A. The Metrics Collector service only reports resource quota usage by user or group once the Quota Management feature is enabled, so its absence explains the missing per-group data even though the service is running.
• B. Restarting the workbench does not activate quota metrics collection, since Prometheus already runs continuously and the missing data is tied to a feature flag, not a stale scan.
• C. The internal NFS filesystem stores project files and code, not resource quota usage data, so adding it as a Grafana data source would not surface quota metrics.
• D. MLAdmin controls who can view Grafana dashboards, not whether quota consumption data is collected for groups in the first place.