Data Management and AI Infrastructure
Data Management and AI Infrastructure
This domain tests the data and platform foundations that make AI repeatable: data quality, lineage, access, contracts, pipelines, feature reuse, and compute capacity. Expect questions on data products, lakehouse patterns, and MLOps infrastructure.
Core rules and concepts
- Treat data as a product: a named owner, documented schema, quality SLOs, and consumers. Data contracts define schema, freshness, and quality between producers and AI teams.
- Assess quality across completeness, accuracy, timeliness, consistency, and label quality. For supervised models, label error is a first-class defect.
- Metadata and lineage answer where data came from, who changed it, and which models depend on it. This is essential for audits and incident response.
- Separate storage, compute, and serving. A lakehouse can support batch and streaming, but AI workloads need capacity planning for GPU and CPU, concurrency, and cost.
- Infrastructure controls: identity and access management, encryption, environment separation, reproducibility, and artifact or version registries.
Worked example
A churn model receives daily customer events. The AI team defines a data contract: event schema v2, freshness by 06:00, null rate under 1%. A pipeline check finds freshness at 09:00 for three days. The model still runs, but features are stale. The team traces lineage to an upstream batch job and blocks promotion until the contract passes. Reasoning: infrastructure should fail fast on contract breaches, not wait for model accuracy to drop weeks later.
Traps
Assuming more data always improves models; ignoring label quality; treating governance as IT-only; no data contracts; and provisioning GPUs without utilization or cost controls.