Section 1: OCI Data Science Introduction and Configuration (20 questions)

1. Which resource is the primary logical container for organizing OCI Data
Science artifacts (projects, notebook sessions, models) within an OCI
tenancy?
A. Region
B. Availability Domain
C. Compartment
D. Fault Domain
Answer: C
Explanation: OCI resources are organized in compartments for access
control and lifecycle management.
2. Which identity construct is best for granting an instance or notebook
session access to Object Storage without embedding credentials?
A. Local user with auth token
B. Dynamic Group with IAM policy
C. API key on a user
D. Resource principal disabled
Answer: B
Explanation: Dynamic groups combined with IAM policies allow
resource principals to access services securely.
3. What is the minimum IAM permission needed for a Data Science user
to create a notebook session in a compartment?
A. inspect
B. read
C. use
D. manage

Data Science
 Answer: D
Explanation: Creating and deleting resources generally requires
manage permissions on the resource family in the compartment.
4. Which statement about OCI regions and Data Science service
availability is correct?
A. Data Science is global-only and not region-scoped
B. Data Science resources are region-scoped; choose the region where
you will run workloads and store data
C. Regions are automatically replicated for Data Science
D. Data Science resources are availability-domain scoped only
Answer: B
Explanation: Data Science resources are created per region and should
be co-located with data for performance.
5. Which network option most securely allows a private notebook session
to access public package repositories?
A. Assign a public IP to the notebook subnet
B. Use an Internet Gateway and route table
C. Use a NAT Gateway from a private subnet
D. Enable local peering
Answer: C
Explanation: Private subnets use NAT Gateway to reach the internet
outbound without exposing public IPs.
6. What is the recommended way to store and access credentials (e.g.,
for private Git or package indexes) from notebook sessions?
A. Hardcode credentials in notebooks
B. Store in Object Storage plaintext
C. Use OCI Vault secrets and retrieve via resource principals
D. Email credentials to team members
Answer: C
Explanation: Vault with resource principals centralizes and secures
secret management.
7. Which resource ties billing and quotas for OCI usage?
A. Tenancy
B. Compartment
C. Project
D. Dynamic Group
Answer: A
Explanation: The tenancy is the root boundary for billing and service
limits.
8. A team wants to isolate dev and prod Data Science resources. What’s
the best-practice approach?
A. Use separate users only
B. Use separate compartments with distinct IAM policies
C. Keep all in root compartment
D. Use tags only
Answer: B

Data Science
 Explanation: Compartmentalization plus policies enforce environment
isolation.
9. To allow a notebook session to read from Object Storage in the same
tenancy, which policy is appropriate?
A. Allow dynamic-group DS-Notebooks to manage object-family in
tenancy
B. Allow dynamic-group DS-Notebooks to read buckets in tenancy
C. Allow group DataScientists to manage object-family in tenancy
D. Allow service objectstorage to manage dynamic-group DS-
Notebooks
Answer: B
Explanation: Least-privilege policy granting read access to buckets to
the notebook dynamic group.
10. Which statement about Data Science projects is accurate?
A. Projects provide network isolation
B. Projects are global across regions
C. Projects group related ML artifacts and metadata within a
compartment
D. Projects replace compartments
Answer: C
Explanation: Projects organize related Data Science resources and
metadata.
11. What does the Data Science Conda environment primarily
control?
A. IAM role bindings
B. VCN routing
C. Runtime packages and versions for notebooks/jobs
D. Object Storage lifecycle policies
Answer: C
Explanation: Conda environments define the software stack used by
notebooks and jobs.
12. How do you prevent unexpected public internet exposure for
Data Science notebook sessions?
A. Place notebooks in a private subnet with no public IP and use NAT
Gateway
B. Place notebooks in a public subnet with Internet Gateway
C. Disable security lists
D. Add default route to [Link]/0 via Service Gateway
Answer: A
Explanation: Private subnet plus NAT allows outbound-only access
without inbound exposure.
13. Where do you enable logging and metrics collection for Data
Science resources?
A. In the tenancy root only
B. Through Logging/Monitoring services integrated with resource-level
settings

Data Science
 C. On the user profile
D. Only via CLI
Answer: B
Explanation: OCI Logging and Monitoring integrate with services and
can be configured per resource/compartment.
14. What is the safest way to grant temporary elevated access to a
data scientist?
A. Promote to tenancy administrator
B. Create a policy using a time-bound group or use approval workflows
with compartments
C. Share API keys
D. Add full access in root forever
Answer: B
Explanation: Principle of least privilege with scoped and time-limited
access.
15. Which SDK is most commonly used for Python-based interaction
with OCI services from notebooks?
A. boto3
B. google-cloud
C. oci
D. azure-identity
Answer: C
Explanation: The OCI Python SDK is the standard for OCI service access
in Python.
16. Which networking component enables private access from a
subnet to Object Storage without using public internet?
A. Internet Gateway
B. NAT Gateway
C. Service Gateway
D. Local Peering Gateway
Answer: C
Explanation: Service Gateway allows private access to OCI services
such as Object Storage.
17. Which statement about compartments and policies is true?
A. Policies can only be written in the root compartment
B. Policies can be attached to any compartment and scoped
accordingly
C. Policies are attached to regions, not compartments
D. Compartments have no effect on authorization
Answer: B
Explanation: Policies can be authored in a compartment and reference
resources and principals across scopes.
18. What is the recommended way to control cost visibility for Data
Science resources?
A. Disable budgets
B. Use cost-tracking tags and Budgets service on

Data Science
 compartments/projects
C. Only rely on monthly invoice
D. Use email reminders manually
Answer: B
Explanation: Budgets plus tags enable granular spend tracking and
alerts.
19. What’s the primary benefit of using resource principals for jobs
and notebook sessions?
A. Avoids need for tenancy-level admins
B. Eliminates IAM completely
C. Provides short-lived, managed credentials scoped by policies
D. Grants unrestricted access to all services
Answer: C
Explanation: Resource principals supply managed auth tokens
constrained by IAM policies.
20. Which is a best practice for organizing ML artifacts in OCI Data
Science?
A. Store models only on local notebook storage
B. Use the Data Science model catalog and Object Storage for
versioned artifacts
C. Email model files between teammates
D. Place all artifacts in root compartment
Answer: B
Explanation: The model catalog and Object Storage support versioning,
lineage, and collaboration.

Section 2: Design and Set Up the OCI Data Science Workspace (30 questions)

1. When creating a private Data Science notebook session, which

networking components are required for outbound internet access
without public exposure?
A. Public subnet + Internet Gateway
B. Private subnet + NAT Gateway + route table
C. Private subnet + Local Peering Gateway
D. Public subnet + Service Gateway
Answer: B
Explanation: A private subnet with a default route to a NAT Gateway
provides outbound-only internet access.
2. What is the primary purpose of a Data Science Project in OCI?
A. Provide VCN-level isolation
B. Group ML resources and metadata (notebooks, jobs, models) within
a compartment
C. Replace compartments for access control

Data Science
 D. Manage billing boundaries
Answer: B
Explanation: Projects organize Data Science artifacts but do not
replace compartments or networking.
3. Which IAM policy grants users the ability to create notebook sessions
in a specific compartment?
A. Allow group DSUsers to read datascience-family in compartment X
B. Allow group DSUsers to use datascience-family in compartment X
C. Allow group DSUsers to manage datascience-family in compartment
X
D. Allow group DSUsers to inspect datascience-family in compartment
X
Answer: C
Explanation: Creating/deleting requires manage level on the Data
Science resource family.
4. What is the best practice for allowing notebooks to access Object
Storage privately?
A. Public subnet + Internet Gateway
B. Service Gateway route to Object Storage
C. Local Peering Gateway only
D. No route; use API keys
Answer: B
Explanation: Service Gateway enables private access to OCI services
like Object Storage.
5. Which construct is best for securing inbound access to a private
notebook session from a jump host?
A. Internet Gateway with open security lists
B. Network Security Groups (NSGs) scoped to the jump host
C. Public IP on the notebook
D. Default security list only
Answer: B
Explanation: NSGs apply fine-grained security rules to instance-level
traffic sources/ports.
6. Which shape selection consideration matters most for GPU-accelerated
training in notebook sessions?
A. OCPU-to-memory ratio only
B. Local NVMe availability only
C. GPU shape family (e.g., A10/A100), GPU count, and VRAM
D. AMD vs Intel CPU only
Answer: C
Explanation: GPU type/count and VRAM drive performance for
accelerated workloads.
7. Where are custom Conda environments for OCI Data Science typically
stored?
A. In the user’s home directory only
B. In Object Storage as conda packs

Data Science
 C. In the root compartment metadata
D. In OCI Vault
Answer: B
Explanation: Custom conda packs are packaged and stored in Object
Storage for reuse across sessions/jobs.
8. To prevent egress to the public internet from a private subnet while
still allowing Object Storage access, use:
A. NAT Gateway
B. Service Gateway only (no NAT route)
C. Internet Gateway
D. DRG
Answer: B
Explanation: Service Gateway provides private access; omitting a NAT
route blocks general internet egress.
9. What is a recommended strategy for isolating dev/test/prod Data
Science resources?
A. Single compartment, use tags to differentiate
B. Separate compartments, policies, and optionally separate VCNs
C. Rely on projects only
D. Store all in root compartment
Answer: B
Explanation: Compartmental isolation plus tailored policies; VCN
isolation as needed.
10. Which artifact defines the code and environment executed by a
Data Science job?
A. Model catalog entry
B. Job and its associated job artifact + conda environment
C. Notebook session only
D. Vault secret
Answer: B
Explanation: Jobs point to a code artifact and an environment (conda)
for execution.
11. How do you enable centralized logging for Data Science jobs and
model deployments?
A. Enable tenancy-wide logging only
B. Attach a Log Group and enable service logs for Data Science
resources
C. Use only print statements
D. Configure logs on user profiles
Answer: B
Explanation: Use OCI Logging with log groups and relevant service log
categories.
12. To allow a notebook session to read/write a private Git repo,
which is most secure?
A. Hardcode SSH key in notebook
B. Store credentials in Object Storage plaintext

Data Science
 C. Use Vault to store secrets and retrieve via resource principals
D. Use public repo only
Answer: C
Explanation: Vault + resource principals provides secure, auditable
secret access.
13. Which is true about Data Science model deployments
networking?
A. Only public endpoints are supported
B. Endpoints can be deployed in a VCN subnet and secured via NSGs
C. No VCN integration is available
D. Requires a load balancer you provision separately
Answer: B
Explanation: Model deployments run in your VCN and can be private or
public depending on subnet/NSG.
14. What’s the advantage of placing data, notebooks, and model
deployments in the same region?
A. Lower tenancy cost
B. Reduced latency and egress costs; simpler IAM and networking
C. Automatic DR
D. Global replication
Answer: B
Explanation: Co-location improves performance and reduces cross-
region egress.
15. Which policy enables notebook sessions (dynamic group) to pull
conda packs from a designated Object Storage bucket?
A. Allow dynamic-group DS-Notes to manage object-family in tenancy
B. Allow dynamic-group DS-Notes to read objects in compartment X
where target bucket resides
C. Allow group DSUsers to manage object-family in tenancy
D. Allow service datascience to manage buckets
Answer: B
Explanation: Least-privilege read access to the bucket for the dynamic
group.
16. What is a best practice for SSH access to private notebook
sessions if direct inbound is disallowed?
A. Assign public IP temporarily
B. Use OCI Bastion with session policies and NSG rules
C. Open [Link]/0 on port 22 in security lists
D. Use cloud shell port-forward to public subnet
Answer: B
Explanation: OCI Bastion provides just-in-time, audited access without
exposing public IPs.
17. How can you constrain Data Science resource costs by shape
selection?
A. Create quotas and limit allowable shapes in compartments
B. Disable users’ notebook creation

Data Science
 C. Use tags only
D. Block all GPU shapes globally
Answer: A
Explanation: Quotas and IAM guardrails control resource types and
sizes per compartment.
18. What is recommended for handling private package indexes
(e.g., Artifactory) from private subnets?
A. NAT Gateway egress + allowlist the index domain
B. Public IP on the subnet
C. No egress allowed
D. Store packages in the notebook filesystem
Answer: A
Explanation: NAT egress preserves privacy while permitting outbound
to approved destinations.
19. What is the role of block storage when provisioning notebook
sessions?
A. Provides persistent user data volume and workspace storage
B. Controls IAM permissions
C. Routes traffic to Service Gateway
D. Manages model versions
Answer: A
Explanation: Block volumes back notebook persistent storage.
20. How do you ensure team-wide reuse of approved conda
environments?
A. Let each user build local environments ad hoc
B. Publish curated conda packs to a central Object Storage bucket and
reference them in projects
C. Use only default environments
D. Store ZIPs on local disks
Answer: B
Explanation: Centralized, versioned conda packs ensure consistency
and compliance.
21. Which is the most secure way to allow a model deployment to
call another OCI service?
A. User-level API keys embedded in code
B. Resource principals with policies via dynamic group for the
deployment
C. Public anonymous access
D. Cross-tenant peering without policies
Answer: B
Explanation: Resource principals for the deployment enable scoped,
managed access.
22. When creating a job to run in a private subnet without internet,
how can you access training data in Object Storage?
A. Internet Gateway
B. NAT Gateway

Data Science
 C. Service Gateway
D. Public IP
Answer: C
Explanation: Service Gateway provides private connectivity to Object
Storage.
23. What is the recommended method to control who can view
model artifacts in the model catalog?
A. Use project settings only
B. Use IAM policies scoped to compartments and datascience-model
resources
C. Rely on tag-based defaults
D. Store models in user home directories
Answer: B
Explanation: IAM policies govern access to model resources.
24. For fine-grained traffic control to a model deployment, which is
preferred?
A. Security Lists only
B. NSGs attached to the deployment and client subnets
C. Open all egress rules
D. No rules needed for private endpoints
Answer: B
Explanation: NSGs provide per-resource ingress/egress control beyond
subnet-wide lists.
25. How do you restrict outbound internet destinations from a
private subnet?
A. Use route tables with NAT plus egress firewall/proxy or restricted
DNS and NSG egress rules
B. Enable Internet Gateway
C. Remove all routes including Service Gateway
D. Use DRG only
Answer: A
Explanation: Combine routing and egress controls to limit destinations.
26. What feature helps track ownership and cost of Data Science
resources?
A. Availability domains
B. Defined tags and cost-tracking tags
C. Fault domains
D. Model catalog metadata only
Answer: B
Explanation: Defined/cost-tracking tags integrate with Budgets and
Cost Analysis.
27. To keep notebook images consistent, you should:
A. Allow users to sudo install system packages freely
B. Use curated conda packs and job/notebook environment specs;
restrict OS-level changes
C. Disable package management

Data Science
 D. Rebuild the OS image daily
Answer: B
Explanation: Curated environments improve reproducibility and reduce
drift.
28. What is the recommended way to connect on-prem data sources
to your Data Science VCN?
A. Public IPs and open security lists
B. Site-to-Site VPN or FastConnect via DRG
C. Internet Gateway only
D. Local peering
Answer: B
Explanation: DRG with VPN/FastConnect provides secure private
connectivity.
29. Which statement about compartment policies and log groups is
correct?
A. Log groups are global and ignore compartments
B. Policies can govern who can create/stream logs and view logs in
specific compartments
C. Logs are always public
D. Log groups require root-compartment policies only
Answer: B
Explanation: IAM policies control Logging operations by compartment.

30. To protect secrets used by jobs and deployments during runtime:

A. Store in environment variables committed to Git
B. Retrieve from OCI Vault at runtime using resource principals and
rotate regularly
C. Store in Object Storage without encryption
D. Encode in Base64
Answer: B
Explanation: Vault with rotation and resource principals is secure and
auditable.

Section 3: Implement the End-to-End Machine Learning Life Cycle (90

questions)

1. Which SDK is the primary toolkit for end-to-end ML workflows on OCI

Data Science?
A. boto3
B. oci
C. ADS (Accelerated Data Science) SDK
D. kubernetes
Answer: C

Data Science
 Explanation: ADS provides high-level APIs for data access, AutoML,
model management, and deployments.
2. What is the most appropriate place to store large training datasets for
use by notebooks and jobs?
A. Local notebook disk only
B. Object Storage
C. Model catalog
D. Vault
Answer: B
Explanation: Object Storage is durable, scalable, and integrates with
Service Gateway for private access.
3. Which ADS capability automates algorithm selection and
hyperparameter tuning?
A. ADS Data Insights
B. ADS AutoML
C. ADS Model Deployment
D. ADS Explainability
Answer: B
Explanation: ADS AutoML automates model search and tuning.
4. For reproducible ML experiments, which combination is best?
A. Manual notes and ad hoc environments
B. Conda packs, job artifacts, and tracked parameters/metrics
C. Only notebook cells
D. Production-only code
Answer: B
Explanation: Versioned environments, code artifacts, and tracked
metadata ensure reproducibility.
5. What’s the recommended way to move from interactive development
to batch training?
A. Copy-paste code into a new notebook
B. Package code as a job artifact and run as an OCI Data Science job
C. Run on local laptop
D. Email scripts
Answer: B
Explanation: Jobs execute packaged code reliably with consistent
environments.
6. Which method helps avoid downloading data over the public internet
from private subnets?
A. Internet Gateway
B. NAT Gateway
C. Service Gateway to Object Storage
D. Public IP on notebook
Answer: C
Explanation: Service Gateway enables private, internal access to
Object Storage.

Data Science
 7. What’s a best practice for tracking experiment parameters and
metrics?
A. Store only in notebook markdown
B. Use ADS tracking or write metrics to Logging/Monitoring/Object
Storage
C. Email CSV logs
D. Ignore metrics during prototyping
Answer: B
Explanation: Persist experiment metadata for comparison and audit.
8. Which is true about ADS Data Insights?
A. Only builds models
B. Provides EDA summaries, schema profiling, and data quality checks
C. Deploys models to endpoints
D. Manages IAM policies
Answer: B
Explanation: Data Insights streamlines EDA and data quality review.
9. To ensure deterministic training runs, you should:
A. Randomize seeds every time
B. Fix random seeds and pin package versions in conda
C. Disable multi-threading always
D. Use any runtime image
Answer: B
Explanation: Fixed seeds and pinned dependencies reduce variance
across runs.
10. Which format is suitable for persisting trained models with
metadata in OCI?
A. Text files
B. Model catalog entries with artifacts in Object Storage
C. Plain local pickle only
D. Email attachments
Answer: B
Explanation: The model catalog tracks versions, artifacts, and lineage.
11. When using GPUs for deep learning training, which is critical?
A. Use CPU-only shapes
B. Ensure CUDA/CuDNN-compatible conda environment
C. Disable virtualization
D. Avoid VRAM usage
Answer: B
Explanation: GPU training needs compatible CUDA/CuDNN libraries.
12. What is the first step in preparing unstructured image data for
model training?
A. Create encryption keys
B. Define dataset schema and labeling conventions
C. Deploy a model
D. Configure NAT only
Answer: B

Data Science
 Explanation: Consistent schema/labels ensure correct ingestion and
training.
13. For large-scale data preprocessing with Spark, which approach
from a notebook is common?
A. Manual loops
B. Use OCI Data Flow (Spark) and read/write via Object Storage
C. Use local-only compute
D. Use Fault Domains
Answer: B
Explanation: Data Flow handles scalable distributed processing with
Object Storage I/O.
14. What’s the recommended way to share common preprocessing
code across notebooks and jobs?
A. Copy/paste into each notebook
B. Package as a Python wheel/ZIP and store as a job artifact
C. Store only in markdown
D. Keep on a local laptop
Answer: B
Explanation: Packaged libraries enable reuse and consistent
versioning.
15. During feature engineering, how should you handle personally
identifiable information (PII)?
A. Store unmasked PII in plain text
B. Apply masking, tokenization, or hashing; restrict access via IAM and
Vault
C. Share via email
D. Ignore governance
Answer: B
Explanation: Protect PII per security and compliance best practices.
16. For hyperparameter tuning at scale on OCI:
A. Manually edit parameters in notebook
B. Use ADS AutoML or orchestrate multiple job runs with parameter
grids
C. Use only defaults
D. Tune in production endpoints
Answer: B
Explanation: AutoML/grid-search via jobs accelerates tuning process.
17. Which is the best place to keep training code and notebooks
under version control?
A. Email chain
B. Private Git repository (e.g., OCI DevOps Code Repo or approved Git)
C. Local filesystem
D. Cloud Console notes
Answer: B
Explanation: Version control enables collaboration and auditability.

Data Science
 18. What’s the purpose of a job run artifact bucket?
A. Store logs only
B. Persist code, dependencies, and outputs for each run
C. Keep IAM policies
D. Manage users
Answer: B
Explanation: Job artifacts centralize inputs/outputs per execution.
19. For class-imbalanced datasets, a common approach is:
A. Ignore the imbalance
B. Use class weights or resampling techniques
C. Reduce features only
D. Increase batch size only
Answer: B
Explanation: Class weights/resampling help address imbalance in
training.
20. Which evaluation metric is most appropriate for binary
classification with imbalanced classes?
A. Accuracy only
B. ROC AUC or PR AUC
C. MSE
D. R2
Answer: B
Explanation: AUC metrics are robust for imbalanced classification.
21. Where should you store intermediate feature datasets for re-use?
A. Local disk of a single notebook
B. Object Storage with versioned prefixes
C. Vault secrets
D. Model catalog exclusively
Answer: B
Explanation: Versioned Object Storage paths support reuse and
lineage.
22. To accelerate repeated dataset reads in training:
A. Disable caching
B. Use parquet format and partitioning in Object Storage
C. Use CSV only
D. Use images embedded in JSON
Answer: B
Explanation: Columnar formats and partitioning improve I/O efficiency.
23. When preparing a job for training, the environment should
include:
A. Only OS packages
B. Exact conda pack or environment spec used in development
C. Random packages
D. No environment specified
Answer: B

Data Science
 Explanation: Matching environments reduce drift between dev and
batch.
24. What’s the advantage of ADS model serialization helpers?
A. Bypass IAM
B. Standardize save/load routines across frameworks
C. Disable encryption
D. Require public networks
Answer: B
Explanation: ADS offers framework-agnostic helpers for persistence.
25. What is a typical step after training a candidate model?
A. Delete it
B. Register it in the model catalog with metadata and evaluation
results
C. Email the binary
D. Deploy immediately without testing
Answer: B
Explanation: Cataloging models captures lineage and promotes
governance.
26. For data validation in pipelines, which technique is common?
A. Skip checks
B. Schema checks, type enforcement, and constraint tests
C. Use only visualizations
D. Randomize schemas
Answer: B
Explanation: Validations prevent bad data from breaking training.
27. What is the purpose of a train/validation/test split?
A. Reduce storage
B. Provide unbiased evaluation and hyperparameter selection
C. Enable public sharing
D. Limit IAM
Answer: B
Explanation: Proper splits avoid overfitting and leakage.
28. To avoid data leakage, you should:
A. Normalize using full dataset statistics
B. Fit transformations on train only, then apply to validation/test
C. Shuffle test into train
D. Use labels to build features
Answer: B
Explanation: Leakage is prevented by isolating test data from fit steps.
29. Which approach speeds up repeated feature transformations?
A. Recompute every run
B. Cache transformed data or persist features
C. Use only CSV
D. Disable parallelism
Answer: B
Explanation: Caching/persisting reduces redundant work.

Data Science
 30. For deep learning frameworks (TensorFlow/PyTorch), which is
best practice on OCI?
A. Use CPU shapes only
B. Use GPU shapes with compatible drivers and pinned framework
versions
C. Avoid VRAM checks
D. Install random nightly builds
Answer: B
Explanation: Compatibility and version pinning ensure stable GPU
usage.
31. Which data splitting strategy preserves class proportions?
A. Random without constraints
B. Stratified split
C. Deterministic hash across features only
D. Time-based split
Answer: B
Explanation: Stratification maintains label distribution across splits.
32. For time-series forecasting, a suitable split is:
A. Random split
B. Shuffle everything
C. Forward-chaining or time-based split
D. Stratified by label
Answer: C
Explanation: Time-aware splits respect temporal order.
33. When running large training jobs, how do you capture logs and
metrics centrally?
A. Only stdout on notebook
B. Stream to OCI Logging/Monitoring or write to Object Storage
C. Print to screen only
D. Use emails
Answer: B
Explanation: Centralized logs and metrics aid troubleshooting and
audit.
34. What is a common way to parameterize jobs?
A. Hardcode constants
B. Use job run arguments/environment variables
C. Use plain comments
D. Use terminal only
Answer: B
Explanation: Parameters allow reusable code across datasets and
experiments.
35. To enable parallel training of multiple models on OCI:
A. Manual loops in one process
B. Launch multiple job runs with different parameter sets
C. Deploy models first
D. Use only one shape

Data Science
 Answer: B
Explanation: Parallel job runs scale grid/AutoML exploration.
36. Which ADS module helps with model explanations?
A. [Link]
B. [Link] (e.g., SHAP)
C. [Link]
D. [Link]
Answer: B
Explanation: ADS provides explainability utilities like SHAP.
37. What is the benefit of storing notebooks and scripts in the
project’s Git repo?
A. Public visibility
B. Versioning, code review, and rollback
C. Automatic deployment
D. Free resources
Answer: B
Explanation: Git supports collaboration and traceability.
38. Which method helps minimize compute costs during EDA?
A. Always use largest GPU shapes
B. Sample data and downscale with lightweight shapes
C. Never sample
D. Run everything on production endpoints
Answer: B
Explanation: Sampling reduces compute time and cost.
39. To ensure confidentiality of training data at rest:
A. Store unencrypted
B. Use default or customer-managed encryption keys for Object
Storage
C. Use public bucket
D. Email archives
Answer: B
Explanation: Object Storage encrypts data, and you can use Vault-
managed keys if required.
40. When building a custom container image for training:
A. Ignore OCI registry
B. Push to OCI Container Registry (OCIR) and reference in jobs
C. Store image in Object Storage
D. Use local Docker only
Answer: B
Explanation: OCIR is the managed container registry for OCI.
41. If training requires both Spark preprocessing and Python model
training:
A. Do both in one notebook only
B. Use Data Flow for Spark stage and a Data Science job for training
C. Use only Spark for training
D. Use only notebooks

Data Science
 Answer: B
Explanation: Combine best-fit services for each stage.
42. How can you ensure your pipeline fails early when input data is
missing?
A. Ignore errors
B. Add pre-check steps that validate data presence and schema
C. Disable logging
D. Skip validation
Answer: B
Explanation: Early checks prevent wasted compute.
43. Which is the safest way to use third-party Python packages in
training?
A. Download from unknown sources
B. Use curated conda packs or private indexes approved by security
C. Install from any random URL
D. Build from unverified forks
Answer: B
Explanation: Follow enterprise-approved sources for compliance.
44. To persist model training seeds and configs:
A. Keep only in memory
B. Store in config files committed to Git and tracked in job metadata
C. Hide in code comments
D. Email them
Answer: B
Explanation: Versioned configs support reproducibility.
45. If a training job needs to read from multiple buckets:
A. Grant manage on object-family in tenancy
B. Grant least-privilege read to the specific buckets for the job’s
dynamic group
C. Use user API keys
D. Make buckets public
Answer: B
Explanation: Principle of least privilege via dynamic group policies.
46. Which is a common pattern for feature normalization?
A. Fit scaler on full dataset
B. Fit on train split; apply to validation/test
C. Fit on test set
D. Don’t normalize
Answer: B
Explanation: Prevents leakage from validation/test.
47. For memory-constrained training on large datasets:
A. Load entire dataset into memory
B. Use chunked streaming, generators, or TFRecord/Parquet
C. Increase swap
D. Ignore OOM errors

Data Science
 Answer: B
Explanation: Streaming and efficient formats avoid OOM.
48. What does ADS opctl primarily help with?
A. IAM policy writing
B. Environment and project bootstrapping from CLI
C. Creating VCNs
D. Budget alerts
Answer: B
Explanation: opctl aids environment setup and project scaffolding.
49. Which approach best documents model lineage?
A. Manual notes only
B. Register models with metadata referencing code repo, data location,
metrics, and environment
C. Store only the binary
D. Keep lineage out of scope
Answer: B
Explanation: Rich metadata supports audit and traceability.
50. How should you store secrets needed during training?
A. Plaintext in notebooks
B. In OCI Vault and fetched via resource principals
C. In Git
D. In [Link]
Answer: B
Explanation: Vault centralizes secure secret management.
51. What is an appropriate validation for text classification pipelines?
A. Skip tokenization checks
B. Verify vocabulary building and sequence length constraints
C. Always pad to zero length
D. Remove stopwords only
Answer: B
Explanation: Proper tokenization/sequence handling is critical.
52. For hyperparameter searches using distributed workers:
A. Use one core only
B. Launch multiple parallel job runs with different params
C. Disable logging
D. Use public endpoints
Answer: B
Explanation: Parallelism speeds up tuning.
53. What is a good practice for handling failed training jobs?
A. Ignore failures
B. Capture error logs, raise clear exceptions, and persist state for
debugging
C. Delete all artifacts
D. Retry blindly
Answer: B
Explanation: Diagnostics facilitate quick remediation.

Data Science
 54. Which is the best metric for regression error magnitude?
A. Accuracy
B. RMSE or MAE
C. ROC AUC
D. F1-score
Answer: B
Explanation: RMSE/MAE quantify regression error.
55. For multiclass classification with uneven classes:
A. Accuracy only
B. Macro-averaged F1
C. R2
D. MAPE
Answer: B
Explanation: Macro F1 treats classes equally regardless of size.
56. To move from exploration to production-ready code:
A. Keep experimental cells
B. Refactor into functions/modules and add tests
C. Leave as-is
D. Only rename notebook
Answer: B
Explanation: Modular, tested code is production-ready.
57. How can you speed up small-to-medium training loops?
A. Use smallest shapes only
B. Utilize vectorized libraries and multi-threading where safe
C. Disable libraries
D. Write pure Python loops
Answer: B
Explanation: Vectorization/parallelism improves performance.
58. Which storage class is cost-effective for infrequently accessed
artifacts (e.g., old checkpoints)?
A. Standard
B. Infrequent Access or Archive tiers
C. Block volume
D. Local NVMe
Answer: B
Explanation: Lower-cost tiers fit infrequent access data.
59. For image model training input pipelines:
A. Load all images into RAM
B. Use batched, prefetching, and on-the-fly augmentation
C. Disable caching
D. Use CSV only
Answer: B
Explanation: Efficient pipelines optimize throughput.
60. What is a correct step before model comparison?
A. Compare on different datasets
B. Evaluate on the same validation/test splits with consistent metrics

Data Science
 C. Use changing metrics
D. Skip evaluation
Answer: B
Explanation: Fair comparison requires consistent data/metrics.
61. Which approach helps track many experiments
programmatically?
A. Manual spreadsheet only
B. Structured metadata in Object Storage or a tracking service via ADS
C. Comments in code
D. Email logs
Answer: B
Explanation: Machine-readable tracking enables queries and
dashboards.
62. For categorical variables with high cardinality:
A. One-hot encode always
B. Consider target/impact encoding or embeddings
C. Drop variable always
D. Use integers as-is
Answer: B
Explanation: Alternative encodings scale better.
63. When should you use cross-validation?
A. When time-series order matters
B. To estimate generalization on limited data (non-time-series)
C. Never
D. Only with test leakage
Answer: B
Explanation: CV improves estimates on small datasets.
64. Which early stopping method is common in gradient boosting?
A. Randomly stop
B. Monitor validation metric and stop after patience rounds
C. Always run max iterations
D. Stop on training loss only
Answer: B
Explanation: Early stopping reduces overfitting and compute time.
65. Which is true about model catalog versions?
A. Only one version allowed
B. Supports multiple versions with metadata and lifecycle states
C. Versions require separate compartments
D. Versions cannot be compared
Answer: B
Explanation: Versioning enables iteration and governance.
66. How do you ensure code that ran in dev will run identically in
jobs?
A. Use different environment
B. Pin and reference the same conda pack/environment spec
C. Upgrade packages mid-run

Data Science
 D. Disable dependency locks
Answer: B
Explanation: Environment parity ensures consistent results.
67. For text models, which preprocessing is commonly needed?
A. Ignore encoding
B. Normalize casing, remove noise, and tokenize
C. Only remove numbers
D. Use raw bytes
Answer: B
Explanation: Standard NLP preprocessing improves model input
quality.
68. To reduce overfitting in neural networks:
A. Increase epochs without regularization
B. Use dropout, weight decay, and data augmentation
C. Use training loss only
D. Increase model size always
Answer: B
Explanation: Regularization and augmentation improve generalization.
69. Which is a correct way to store large training logs?
A. In notebook memory
B. Stream to Object Storage or Logging
C. Only stdout
D. Email text
Answer: B
Explanation: External storage avoids memory bloat and aids analysis.
70. When validating data joins:
A. Assume keys match
B. Check for duplicates, nulls, and referential integrity
C. Ignore nulls
D. Drop keys
Answer: B
Explanation: Join validation prevents silent data corruption.
71. For distributed hyperparameter tuning that may fail partially:
A. Fail the entire batch on first error
B. Make runs independent and aggregate best results afterward
C. Stop all on warning
D. Use single-thread only
Answer: B
Explanation: Independence improves robustness and throughput.
72. After selecting a champion model, next step is to:
A. Delete training data
B. Register with full lineage and add evaluation artifacts
C. Disable logging
D. Ignore metrics
Answer: B
Explanation: Registration preserves traceability for promotion.

Data Science
 73. To run data preprocessing nightly:
A. Manually trigger notebooks
B. Create a Data Science job and schedule via external scheduler or
OCI Functions/Events
C. Wait for users
D. Use only interactive runs
Answer: B
Explanation: Jobs enable reliable, automated batch processing.
74. For binary classification threshold selection:
A. Always 0.5
B. Optimize threshold based on business objective (e.g., F1,
precision@k)
C. Random threshold
D. Highest recall only
Answer: B
Explanation: Thresholds should align with cost/benefit trade-offs.
75. Which helps diagnose training instability?
A. Ignore seeds
B. Log random seeds, learning rates, loss curves, and system info
C. Hide logs
D. Only final metrics
Answer: B
Explanation: Detailed diagnostics reveal instability causes.
76. For scalable feature computation on tabular data:
A. Use single-threaded loops
B. Use vectorized Pandas/NumPy or Spark when needed
C. Use plain text parsing
D. Only Python lists
Answer: B
Explanation: Vectorization and distributed compute scale better.
77. To minimize costs for hyperparameter tuning:
A. Use largest shapes
B. Use small shapes for broad search, larger only for finalists
C. Always GPUs
D. Unlimited parallel runs
Answer: B
Explanation: Stage resources based on search phase.
78. Which artifact should be versioned with the model?
A. Only weights
B. Training code commit hash, data reference, environment, metrics,
and preprocessing pipeline
C. Only metrics
D. Only data
Answer: B
Explanation: Full context enables reproducibility.

Data Science
 79. How do you persist pandas profiling or EDA summaries?
A. Keep in RAM
B. Save HTML/JSON to Object Storage with run ID
C. Paste into chat
D. Drop after view
Answer: B
Explanation: Persisted EDA supports review and audit.
80. For large CSVs in Object Storage:
A. Read without compression
B. Prefer Parquet or compressed formats for efficiency
C. Use XLSX
D. Split randomly
Answer: B
Explanation: Columnar/compressed formats reduce I/O and cost.
81. To avoid mixing dev/test credentials:
A. Hardcode creds
B. Use separate Vault secrets and IAM policies per environment
C. Share tokens across envs
D. Store in notebooks
Answer: B
Explanation: Environment isolation prevents cross-contamination.
82. For model selection with many candidates:
A. Choose highest training score
B. Compare on held-out validation/test with consistent metrics and
fairness checks
C. Use random choice
D. Choose smallest file
Answer: B
Explanation: Selection must rely on unbiased evaluation.
83. What is a good practice for handling random state across
libraries?
A. Ignore
B. Set seeds for NumPy, Python, and framework-specific RNGs
C. Only framework seed
D. Only OS seed
Answer: B
Explanation: Multiple RNGs require coordinated seeding.
84. For schema evolution in input data:
A. Fail silently
B. Implement schema versioning and backward-compatible parsers
C. Ignore new fields
D. Drop schema
Answer: B
Explanation: Versioning helps manage changes safely.
85. To capture hardware context of training runs:
A. Not needed

Data Science
 B. Record shape, OCPUs, memory, GPU type/count, and driver versions
C. Only note region
D. Only log CPU
Answer: B
Explanation: Hardware context affects performance and
reproducibility.
86. For probabilistic predictions, how do you calibrate outputs?
A. Ignore calibration
B. Use Platt scaling or isotonic regression on validation data
C. Use threshold only
D. Use training set
Answer: B
Explanation: Calibration adjusts predicted probabilities to reflect
reality.
87. When exporting trained models for external scoring:
A. Use proprietary format only
B. Export in common formats (e.g., ONNX where feasible) alongside
environment specs
C. Only pickle
D. Email files
Answer: B
Explanation: Interoperable formats and env specs ease portability.
88. To prevent accidental data mix-ups across runs:
A. Reuse same path
B. Use unique, run-scoped prefixes in Object Storage
C. Overwrite previous outputs
D. Use local temp only
Answer: B
Explanation: Run-specific paths avoid collisions.
89. For labeling quality assessment in supervised learning:
A. Trust labels blindly
B. Sample and audit labels; compute inter-annotator agreement
C. Remove labels
D. Use test labels for training
Answer: B
Explanation: Label audits ensure training signal quality.
90. Before handing off a model for deployment:
A. Skip documentation
B. Provide a model card with purpose, data, metrics, limitations, and
risk considerations
C. Only give binary
D. Change metrics
Answer: B
Explanation: Model cards document context and responsible use.

Data Science
 Section 4: Apply MLOps Practices (40 questions)

1. What is the standard way to serve models in OCI Data Science?

A. Use OKE only
B. Data Science Model Deployment endpoints
C. Compute instance with custom Flask
D. API Gateway alone
Answer: B
Explanation: Model deployments are managed endpoints with scaling,
logging, and VCN integration.
2. Which artifact is required to create a model deployment?
A. Only a notebook
B. A registered model artifact and an inference environment (conda
pack/spec)
C. Vault secret only
D. Data Flow application
Answer: B
Explanation: Deployments reference a cataloged model plus an
environment for inference.
3. To restrict access to a private model endpoint, you should:
A. Make it public
B. Place it in a private subnet and control via NSGs and IAM
C. Open all ports in security lists
D. Use only tags
Answer: B
Explanation: Private subnet + NSGs + IAM policies enforce access
control.
4. How do you enable canary-style releases for model endpoints?
A. Randomly redeploy
B. Run parallel deployments (old/new) and shift traffic using API
Gateway stage or client routing
C. Delete old endpoint
D. Use only DNS caching
Answer: B
Explanation: Parallel endpoints with controlled routing enable gradual
rollout.
5. What’s a sensible health check for model deployments?
A. No health checks
B. Lightweight /health endpoint to validate environment readiness
C. Full training run
D. Email ping
Answer: B
Explanation: Simple readiness checks detect startup issues.
6. How should you handle secrets in inference code?
A. Hardcode in code
B. Retrieve from OCI Vault via resource principals at runtime

Data Science
 C. Store in plaintext env vars
D. Ship in model artifact
Answer: B
Explanation: Vault with resource principals is secure and auditable.
7. For autoscaling inference capacity, which approach is valid?
A. Change tenancy limits
B. Adjust model deployment instance count via API/terraform and
monitor utilization
C. Use only one replica
D. Add more compartments
Answer: B
Explanation: Scale replicas programmatically based on metrics.
8. To capture inference request/response logs with PII control:
A. Log everything
B. Use structured logs, mask/redact PII, send to OCI Logging
C. No logging in prod
D. Store in public bucket
Answer: B
Explanation: Structured, redacted logs meet compliance.
9. How do you detect data drift in production?
A. Ignore
B. Compare live feature distributions to training baselines; log metrics
and set alarms
C. Use only accuracy
D. Wait for outages
Answer: B
Explanation: Drift metrics trigger proactive retraining or investigation.
10. What is a safe deployment strategy to minimize downtime?
A. In-place replace only
B. Blue/green with switch-over via API Gateway or client config
C. Delete and recreate
D. Nightly random redeploy
Answer: B
Explanation: Blue/green avoids outages and eases rollback.
11. Which service helps trigger retraining when new data lands in a
bucket?
A. Budgets
B. Events service invoking Functions to start a Data Science job
C. Audit
D. Bastion
Answer: B
Explanation: Events + Functions orchestrate automated job runs.
12. What’s the best way to store model artifacts for deployment?
A. Local disk
B. Model catalog with versioning; artifacts in Object Storage
C. Plain email

Data Science
 D. Untracked VM folder
Answer: B
Explanation: Catalog + Object Storage ensures traceability.
13. How do you enforce least-privilege for deployments accessing
other services?
A. Admin access
B. Dynamic group for deployments with narrowly scoped IAM policies
C. User API keys in code
D. Public buckets
Answer: B
Explanation: Resource principals scoped via policies.
14. A/B testing of two model versions is best achieved by:
A. Random client selection across two endpoints and aggregate
outcomes
B. Always newest model only
C. Manual handoffs
D. Ignore metrics
Answer: A
Explanation: Split traffic and compare business metrics.
15. Which metric is most useful for scaling inference replicas?
A. Number of compartments
B. CPU/GPU utilization, latency percentiles, request rate (RPS)
C. Region count
D. Budget limit
Answer: B
Explanation: Load and latency drive capacity needs.
16. Where should you push custom inference containers?
A. Local Docker only
B. OCI Container Registry (OCIR)
C. Public GitHub releases
D. Compute instance disk
Answer: B
Explanation: OCIR is the managed registry for OCI.
17. To secure internet-facing endpoints:
A. No front door
B. Front with API Gateway and WAF for auth, throttling, and protection
C. Public subnet with open ports
D. Rely on obscurity
Answer: B
Explanation: API Gateway + WAF adds auth and layer-7 controls.
18. How should you handle model rollback?
A. Delete history
B. Keep previous model deployment ready and switch traffic back
C. Rebuild from memory
D. Block changes

Data Science
 Answer: B
Explanation: Fast rollback reduces MTTR.
19. What’s the role of CI/CD in MLOps on OCI?
A. None
B. Build, test, and deliver model/code artifacts via OCI DevOps
pipelines
C. Only manual deploys
D. Only code review
Answer: B
Explanation: OCI DevOps automates build/test/deploy.
20. For governance and audit of deployments:
A. Skip audit
B. Use Audit service, Logging, and model catalog lineage
C. Email approvals
D. Store in notebooks
Answer: B
Explanation: Native services provide audit trails.
21. How can you expose a private model endpoint to on-prem clients
securely?
A. Public IP
B. FastConnect or VPN via DRG to the VCN
C. NAT Gateway inbound
D. SSH tunnel only
Answer: B
Explanation: DRG with private connectivity is secure and reliable.
22. To avoid breaking changes in scoring:
A. Change schemas silently
B. Version your input/output schemas and keep backward-compatible
handlers
C. Ignore schema changes
D. Force clients to upgrade instantly
Answer: B
Explanation: Schema versioning protects clients.
23. Where to store model performance SLOs and alerts?
A. In comments only
B. Monitoring metrics and Alarms; document SLOs with runbooks
C. Local text files
D. None
Answer: B
Explanation: Metrics + Alarms operationalize SLOs.
24. How should you load large models efficiently at startup?
A. Synchronous blocking only
B. Lazy load where possible, warm-up requests, cache models
C. Reload per request
D. Avoid caching

Data Science
 Answer: B
Explanation: Warm-up and caching improve latency.
25. To protect endpoints from abuse:
A. Unlimited requests
B. Throttling and quotas via API Gateway/WAF; auth with IAM/IDCS/JWT
C. Open traffic
D. No auth
Answer: B
Explanation: Rate limits and auth mitigate abuse.
26. For environment drift prevention in deployments:
A. Allow ad hoc pip installs
B. Pin conda packs; immutable versions; rebuild image on change
C. Install nightly packages
D. Use latest always
Answer: B
Explanation: Pinned environments ensure consistency.
27. Which practice helps quantify model fairness in production?
A. Ignore segments
B. Monitor metrics by cohort/segment and compare to baselines
C. Only global accuracy
D. Manual anecdotes
Answer: B
Explanation: Segment monitoring detects disparate impact.
28. How to store runtime feature logs for drift analysis cost-
effectively?
A. Keep only in memory
B. Write batched JSON/Parquet to Object Storage with lifecycle policies
C. Put in Vault
D. Email CSVs
Answer: B
Explanation: Object Storage is durable and inexpensive.
29. How do you validate a new model before full traffic?
A. Full cutover immediately
B. Shadow traffic: send copies of requests to the new model, compare
offline
C. Ignore validation
D. Only unit tests
Answer: B
Explanation: Shadowing reduces risk pre-cutover.
30. What’s recommended for secrets rotation affecting
deployments?
A. Manual reuse forever
B. Use Vault rotation and ensure deployments fetch secrets at startup
or refresh on signal
C. Embed static keys
D. Restart monthly manually only

Data Science
 Answer: B
Explanation: Automated rotation ensures continuous security.
31. To standardize deployment creation:
A. Manual console clicks only
B. Use IaC (Terraform/Resource Manager) modules for repeatable
deployments
C. Store steps in a doc
D. Shell history
Answer: B
Explanation: IaC ensures consistent, auditable provisioning.
32. How to test inference code during CI?
A. Skip tests
B. Run unit and contract tests with mocked models and sample
payloads
C. Only UI tests
D. Only load tests
Answer: B
Explanation: Tests catch defects early.
33. What is the safest credential type for a calling service-to-service
client?
A. Static user API keys
B. Instance principal/resource principal with IAM policies
C. Basic auth
D. Shared token via email
Answer: B
Explanation: Principals provide short-lived, managed auth.
34. Which log type is most helpful for troubleshooting cold start
issues?
A. Only access logs
B. Application logs during model load and init
C. Audit logs only
D. None
Answer: B
Explanation: Startup logs reveal load/init problems.
35. For blue/green infra changes to VCN or subnets:
A. Live edit production
B. Provision parallel infra and switch endpoints after validation
C. Destroy and recreate
D. Skip testing
Answer: B
Explanation: Parallel infra reduces risk.
36. How to avoid model file tampering in transit to the deployment?
A. HTTP
B. Use signed, encrypted Object Storage and verify checksums/hashes
C. FTP
D. Email

Data Science
 Answer: B
Explanation: Encryption and integrity checks protect artifacts.
37. What helps enforce org-wide deployment standards?
A. Tribal knowledge
B. Golden templates/modules (Terraform), defined tags, and policy
guardrails
C. Per-user freestyle
D. Ad hoc reviews
Answer: B
Explanation: Templates and policies drive consistency.
38. To monitor end-to-end latency including gateway and backend:
A. Backend logs only
B. Collect metrics at API Gateway and inside model deployment;
correlate with request IDs
C. Ignore gateway
D. Only average latency
Answer: B
Explanation: Multi-hop metrics provide full picture.
39. What’s a recommended rollback trigger?
A. Developer intuition
B. Alarms on error rate/latency SLO breaches with automated revert
C. Weekly meetings
D. Customer tweets
Answer: B
Explanation: SLO breaches should trigger automated rollback.
40. How to handle compliance evidence for models in production?
A. Not needed
B. Maintain model cards, approvals, lineage, test reports, and
deployment change records
C. Only code comments
D. Delete history
Answer: B
Explanation: Documented evidence supports audits and compliance.

Section 5: Use Related OCI Services (20 questions)

1. Which service provides managed Spark for large-scale data

processing?
A. Data Flow
B. Functions
C. API Gateway
D. Bastion
Answer: A
Explanation: Data Flow runs serverless Spark with Object Storage I/O.
2. What service stores and rotates encryption keys and secrets?
A. Logging

Data Science
 B. Vault
C. Monitoring
D. Budgets
Answer: B
Explanation: OCI Vault manages keys and secrets securely.
3. Which service should front an internet-facing inference API?
A. Data Catalog
B. API Gateway (optionally with WAF)
C. Streaming
D. Resource Manager
Answer: B
Explanation: API Gateway manages routing, auth, and throttling.
4. To schedule and automate retraining with events, use:
A. Budgets
B. Events + Functions to trigger Data Science jobs
C. DNS
D. Bastion
Answer: B
Explanation: Event-driven functions orchestrate job runs.
5. Which service is best for storing structured model scoring logs at
scale?
A. File system only
B. Logging service with Log Groups and custom logs
C. Vault
D. Audit only
Answer: B
Explanation: Logging provides scalable, searchable log ingestion.
6. For real-time feature ingestion and streaming predictions:
A. Budgets
B. OCI Streaming with consumer apps; store features in Object
Storage/ADB
C. DNS
D. Email
Answer: B
Explanation: Streaming enables low-latency data pipelines.
7. Where should Docker images for training/inference be stored?
A. Local workstation
B. OCIR
C. Public repo only
D. Compute boot volume
Answer: B
Explanation: OCIR is the managed container registry.
8. Which service helps catalog and govern data assets used in ML?
A. Data Catalog
B. DevOps
C. Events

Data Science
 D. Bastion
Answer: A
Explanation: Data Catalog provides metadata and governance
integration.
9. How do you implement IaC for ML infrastructure on OCI?
A. Manual clicks
B. Terraform via OCI Resource Manager
C. Only shell scripts
D. Local spreadsheets
Answer: B
Explanation: Resource Manager executes Terraform in OCI.
10. For private connectivity between on-prem and your ML VCN:
A. Internet Gateway
B. DRG with Site-to-Site VPN or FastConnect
C. NAT only
D. Public peering
Answer: B
Explanation: DRG provides private connectivity options.
11. Which service enables CI/CD for ML code and artifacts?
A. DNS
B. DevOps (Code Repo, Build, Deploy)
C. Data Safe
D. Notifications only
Answer: B
Explanation: OCI DevOps supports repositories and pipelines.
12. Where should training datasets be stored for durability and cost
efficiency?
A. Local NVMe only
B. Object Storage Standard/IA tiers
C. Boot volumes
D. Email attachments
Answer: B
Explanation: Object Storage is durable and cost-effective.
13. To notify teams on job failures:
A. Manual emails
B. Notifications service integrated with Alarms/Events to Slack/Email
C. Audit logs only
D. No alerts
Answer: B
Explanation: Notifications distributes alerts to channels.
14. Which database is commonly used for features or metadata with
SQL access?
A. Text files
B. Autonomous Database (ATP/ADW)
C. Local SQLite only
D. Email

Data Science
 Answer: B
Explanation: Autonomous Database provides managed, scalable SQL.
15. For secure admin access to private subnets:
A. Public SSH
B. OCI Bastion sessions
C. Open [Link]/0
D. RDP exposed
Answer: B
Explanation: Bastion offers JIT, audited access.
16. Which service provides metrics and alarms for infrastructure and
applications?
A. Logging
B. Monitoring
C. Audit
D. Data Catalog
Answer: B
Explanation: Monitoring stores metrics and triggers alarms.
17. What is best for serverless model post-processing or
orchestration glue?
A. Functions
B. Data Flow
C. OKE
D. DNS
Answer: A
Explanation: Functions implement lightweight serverless logic.
18. Which service gives an immutable record of API operations in
OCI?
A. Monitoring
B. Audit
C. Notifications
D. Budgets
Answer: B
Explanation: Audit logs capture API calls for governance and forensics.
19. For hosting a custom microservice that wraps ML with additional
logic at scale:
A. Functions only
B. OKE (Kubernetes) or Compute with Load Balancer
C. Data Flow
D. API Gateway alone
Answer: B
Explanation: OKE/Compute supports long-running services and
autoscaling.
20. How can you protect public-facing APIs from common web
threats and do rate limiting?
A. Security lists only
B. WAF in front of API Gateway or Load Balancer

Data Science
 C. DRG
D. Bastion
Answer: B
Explanation: WAF adds OWASP protections and rate controls.

Data Science