Meridian's predictive-maintenance model is trained monthly on a large historical archive of sensor readings and maintenance outcomes, but in production it must score a live stream of current sensor readings within seconds to flag an at-risk component. What should the data engineering team build to support both needs?

Training and inference have fundamentally different data needs: training processes large historical batches with no strict latency requirement, while inference must score live or near-live data within tight latency bounds. Building separate, purpose-optimized pipelines for each is a named CPMAI Data for AI enabler precisely because a single undifferentiated pipeline usually fails one of the two use cases. Option A ignores this mismatch and risks either slow inference (if built like a batch training pipeline) or inadequate historical processing (if built like a lean inference pipeline). Option C is wrong because relying on stale cached data for inference would mean the model scores against outdated readings, defeating the goal of catching an at-risk component quickly. Option D wrongly assumes historical data becomes irrelevant post-deployment, when in fact it remains essential for retraining, drift monitoring, and model iteration. For the exam, this training/inference pipeline split is one of the most concretely testable Data for AI facts — expect scenario questions that hinge on recognizing which pipeline a described data flow belongs to.