Develop end-to-end ML pipelines that validate data and models, orchestrate managed and unmanaged services from templates or custom solutions such as Agent Platform Pipelines, Managed Service for Apache Airflow, and Ray on the Agent Platform, and ensure consistent data preprocessing between training and serving.

A fraud team serves a model behind a low-latency online endpoint. During training a Vertex AI Pipeline computes categorical encodings and normalisation statistics from the full training set and writes them to a feature store. The online service currently recomputes those same statistics on the fly from each incoming batch of requests before predicting. Reviewers want to guarantee that the values fed to the model at serving time match those used during training. What change achieves this with the least risk of skew?

• AHave the online service read the precomputed encodings and normalisation statistics from the feature store written by the pipeline, instead of recomputing them per request.check_circle Correct
• BIncrease the request batch size at serving so the per-batch statistics the service computes converge closer to the training-set statistics over time.
• CRetrain the model more frequently so its learned weights stay aligned with whatever statistics the serving service happens to compute from recent traffic.
• DCache each request's computed statistics in the service and reuse them for subsequent requests to keep the serving values stable across calls.

Why A is correct: Reading the exact statistics the pipeline persisted guarantees serving applies the same encodings and normalisation as training, eliminating skew while also removing per-request recomputation cost.

Why B is wrong: Larger batches make per-request estimates less noisy, which seems to narrow the gap, but online traffic is not distributed like the training set so the recomputed values never match the persisted ones and skew persists.

Why C is wrong: Frequent retraining is a real defence against drift and therefore tempting, but it chases a moving serving-side computation rather than fixing it, leaving training and serving on different transformation logic between retrains.

Why D is wrong: Caching stabilises serving values call to call and looks like consistency, but the cached values still originate from live traffic rather than the training set, so they remain misaligned with the statistics the model was trained on.