Data Classification and Sensitivity Handling Questions

Classifying data by sensitivity and applying controls proportionate to that classification: identifying personal, sensitive, and special-category data and tagging it through its lifecycle. Covers classification schemes, labeling, and how classification drives access, encryption, and retention decisions. Includes assessing the impact of a given data type on privacy and security risk.

HardSystem Design

38 practiced

Design a data governance model for training data that balances broad access for research and tight controls for privacy, lineage, and quality. Describe policies, access controls, metadata, tooling, and enforcement mechanisms that scale across teams.

Sample Answer

Requirements & constraints- Broad research access for experimentation and reproducibility- Strong privacy (PII masking, consent), data lineage, quality guarantees- Scales across teams, cloud regions, and model sizes- Enforceable, auditable, low friction for researchers

High-level architecture- Raw data lake (immutable) → Ingest & preprocessing layer → Curated training datasets (versioned) → Model training sandboxes → Model registry / artifacts- Central metadata & policy plane + enforcement sidecars per compute cluster

Policies (examples)- Classification policy: Data tagged (public, internal, sensitive, restricted) with required controls per tag- Privacy policy: PII detection → require tokenization, pseudonymization, or differential privacy for sensitive tags before training- Lineage & reproducibility policy: Every dataset and model must have immutable provenance, manifest, and data version referenced- Access & usage policy: Purpose-based access (research, production), expiry for demo access, and allowed compute tiers

Access controls & enforcement- Authentication: SSO + MFA; enterprise IdP- Authorization: Combined RBAC + ABAC (role + dataset attributes + purpose). Example: researcher role + purpose=“explore” grants read to synthetic and masked datasets; purpose=“train” requires approval for sensitive datasets- Fine-grained data plane: Column-level encryption and tokenization; row-level filters enforced via query gateway (e.g., Spark/Presto sidecar)- Policy engine: Open Policy Agent (OPA)/Gatekeeper evaluate policies at request time- Entitlement workflows: Automated approvals (Jira/ServiceNow) for exceptions with TTL

Metadata & lineage- Central catalog: DataHub/Amundsen capturing schema, tags, owners, sensitivity, data contracts, quality scores- Lineage: Instrument ingestion, transformation, training jobs to emit provenance (dataset version, code commit, hyperparams) to catalog; store immutable manifests (hashes)- Versioning: Use object store versioning + LakeFS/Delta/Zarr for dataset snapshots; DVC or MLFlow for model-data links

Quality & validation tooling- Data contracts + expectations: Great Expectations or custom validators run on ingest and pre-train; fail gating for downstream training- Continuous monitoring: Data drift detectors, statistical tests; automated alerts and rollbacks- Test-integration: Pre-train CI that replays a subset with known labels to validate quality

Privacy-preserving tooling- PII discovery: Automated scanners (regex, ML detectors) and manual tagging- Transformation library: Standardized, audited transformations—redaction, hashing, tokenization, k-anonymity, differential privacy (DP-SGD library)- Synthetic data generator: For high-risk datasets provide synthetic substitutes with fidelity metrics

Operational & scaling considerations- Enforcement sidecars co-located with compute to avoid central bottleneck; policy cache for low-latency decisions- Multi-tenant quotas and compute isolation (namespaces, IAM roles)- Auditability: Immutable audit logs to SIEM (append-only, signed) for every access, transformation, and model training run- Automation: Templates for dataset provisioning, policy application, and approval flows to minimize friction- Cost & performance: Tiered data storage (hot for active experiments, cold for archive); sample+proxy datasets for quick prototyping

Governance lifecycle & org processes- Data steward role per domain: approves classification, owners respond to drift alerts- Periodic reviews: Sensitivity reclassification, policy tuning, and model risk assessments- Onboarding & training: Self-service docs, policy-as-code examples, and “safe playground” datasets

Metrics to measure success- Time-to-provision safe dataset- % of training runs referencing versioned datasets- Number of policy violations detected/blocked- Drift incidents and time-to-remediation- Audit completeness and mean time to reconcile access requests

Trade-offs- Strict controls add friction => mitigate with automation, templated approvals, and synthetic/sampled datasets- DP reduces utility — use for highest-risk data and pair with synthetic augmentation

This design balances researcher velocity with privacy, lineage, and quality by combining metadata-driven policies, enforcement at the data plane, automated validation, and audited provenance that scale across teams.

Unlock Full Question Bank

Get access to hundreds of Data Classification and Sensitivity Handling interview questions and detailed answers.

Join thousands of developers preparing for their dream job.