Vertex AI and the ML Platform

Vertex AI brings the stages of a machine learning workflow into one managed platform: preparing data, training models, evaluating them, deploying them for prediction, orchestrating the whole thing as a pipeline, and monitoring what happens afterwards.

It supports several levels of involvement — automated training that handles model selection for you, custom training where you supply the code, and access to pre-trained foundation models you call directly.

The difficulty in machine learning is rarely training a model that works in a notebook. It is everything after: reproducing the result, serving it at acceptable latency, keeping training and serving features consistent, retraining when data shifts, and noticing when quality degrades.

This is the MLOps problem, and it is where the ML Engineer exam spends most of its attention. Treating it as a modelling exam is the most common preparation mistake.

• Training job — running your training code on managed infrastructure, optionally distributed across machines or accelerators.
• Automated training — the platform handles model selection and tuning from a dataset. A strong baseline and often sufficient for tabular problems.
• Model registry — versioned storage for trained models, with the metadata needed to know how each was produced.
• Endpoint — a deployed model serving online predictions, with autoscaling and traffic splitting between versions.
• Batch prediction — scoring a large dataset offline, without a persistent endpoint. Cheaper when latency does not matter.
• Feature store — a central place for features, so training and serving compute them identically. This is what prevents training–serving skew.
• Pipelines — orchestrated, reproducible workflows chaining data preparation, training, evaluation and deployment.
• Model monitoring — watching production inputs and predictions for drift and skew, triggering alerts or retraining.
• Explainability — attributing a prediction to its contributing features, which is often a regulatory requirement.

• Training a custom model on data too large for a single machine.
• Serving low-latency predictions inside a user-facing request.
• Scoring a large dataset in batch on a nightly schedule.
• Automating retraining when monitoring detects drift.
• Sharing consistent features across several models and teams.
• Building on foundation models with tuning and grounding rather than training from scratch.