Oracle Cloud Infrastructure 2025 Data Science Professional Advanced Practice Exam: Hard Questions 2025

You operate an OCI Data Science project where multiple teams use the same VCN. A notebook session must access a private Object Storage bucket via a Service Gateway, call an internal model endpoint deployed behind a private load balancer, and fetch secrets from Vault. The notebook must not have public internet access, and access must be restricted so that only this notebook session can read the bucket objects and secrets. Which design best meets the requirement with least privilege and correct OCI networking behavior?

A Data Science team uses jobs for nightly feature generation. The job runs in a private subnet and must write logs/metrics for troubleshooting and auditing. After tightening network rules, the job starts failing with timeouts when attempting to write logs. There is no Internet or NAT Gateway by design. Which change most directly restores logging while preserving the no-internet constraint?

Your organization mandates that training artifacts and intermediate datasets never leave a specific OCI compartment. Data scientists run notebook sessions that automatically create and update model artifacts and save them to Object Storage. During an audit, you discover that some runs wrote artifacts to a bucket in a different compartment due to default configuration in code. What is the most robust OCI-native control to prevent future cross-compartment writes while still allowing notebooks to read shared reference data from a central bucket?

You are building a fraud model with highly imbalanced classes. The dataset is time-ordered and exhibits concept drift. A junior engineer proposes random train/test split and using accuracy as the primary metric. You must produce an evaluation protocol that is robust to leakage and drift, and that aligns with the business goal of minimizing false negatives while controlling false positives. Which approach is most appropriate?

A notebook pipeline trains a gradient-boosted model using target encoding for high-cardinality categorical features. In production, the deployed model’s performance drops sharply and exhibits suspiciously optimistic offline validation results. After investigation, you find the encoding was fit on the full dataset before splitting into train/validation. What is the best corrective action to prevent this class of leakage going forward while keeping the workflow reproducible in OCI Data Science?

You train a model that uses both numeric and text features. The text vectorizer is stochastic due to hashing and the training uses multi-threaded execution. You must meet an audit requirement: given the same input data snapshot, you must be able to reproduce the exact same model artifact bit-for-bit. Which set of actions is most appropriate to maximize determinism in OCI Data Science training runs?

You deploy a model as an OCI Data Science model deployment behind a private endpoint. Under load testing, p95 latency spikes and occasional 502 errors appear, but CPU utilization stays low. Logs show long pauses during model initialization for some replicas. You need to reduce cold-start effects and stabilize latency without overprovisioning. Which change is most likely to address the root cause?

A regulated enterprise requires that any model deployed to production must be approved, immutable, and fully traceable to training data, code, and evaluation results. Teams currently deploy by directly pointing deployments to the “latest” object in a bucket. Auditors found that re-uploading an object changed the production model without a formal approval event. Which design most effectively enforces immutability and approval gates in OCI?

A batch scoring job uses an OCI Data Science job to read features from Object Storage and write predictions back. After a security hardening, the job intermittently fails with “NotAuthorizedOrNotFound” when reading objects. The same code works from a developer laptop using API keys. The job runs with resource principal. Which troubleshooting path is most likely to identify and fix the issue?

You need an end-to-end MLOps design on OCI where feature generation is performed daily, training occurs weekly, and deployments happen only after automated tests pass. Requirements: (1) orchestrate steps with clear lineage, (2) minimize data movement, (3) support rollback to a previous production model, and (4) ensure least-privilege access between steps using OCI-native identity. Which architecture best satisfies these constraints?