Prepare and test your skills
Prepare and test your skills
Manage Model Deployment and Versioning Post-Training is the process of moving a trained AutoML model from training into production inference while keeping every version traceable and reversible. On Gemini Enterprise Agent Platform (opens in a new tab), the transition flows from the AutoML TrainingPipeline resource to Model Registry, then to an endpoint or batch prediction service, with monitoring attached after deployment. Versioning covers code, data, features, and model binaries together so a deployed version can be reproduced, compared, and rolled back.
The transition from training to inference starts when the AutoML TrainingPipeline (opens in a new tab) finishes training and produces a versioned model for deployment. The TrainingPipeline resource always executes the training task and can also export data from a Gemini Enterprise Agent Platform Dataset as training input, upload the resulting model to the platform, and evaluate the model. Uploaded models become deployment candidates that flow from evaluation to a serving target for online endpoints or batch prediction, which lets training, evaluation, and serving evolve independently while preserving lineage from dataset to deployed model.
The handoff follows an ordered sequence managed by Agent Platform Pipelines (opens in a new tab):
Dataset into the TrainingPipeline, which runs AutoML training.Version control across code, data, and features means each artifact type is stored in a system that supports versioning and promotion from development to production. Code versions are managed in GitHub or GitLab with branching strategies and integration with continuous integration and continuous delivery (CI/CD) pipelines. Data, feature, and container artifacts use Google Cloud storage services so the exact inputs behind a model version can be recovered for retraining or audit.
The artifact-to-service mapping defines ownership and dependencies for reproducibility:
| Artifact type | Service | Versioning role |
|---|---|---|
| Container images and pipeline components | Artifact Registry | Versions code artifacts, improves security, and integrates with Cloud Build and Cloud Deploy |
| Training and evaluation datasets | BigQuery or Cloud Storage | Stores and versions data artifacts and supports large-scale analysis |
| Feature data | Agent Platform Feature Store | Maintains consistency and versioning of feature data with BigQuery as the offline store |
| Metadata and data pointers | Version control system or data catalog | Stores locations and links that connect code, data, and model versions |
Feature consistency depends on Feature Store, which serves low-latency workloads from feature views backed by BigQuery tables or views. Because BigQuery retains historical feature data for offline use, a deployed model version can be joined back to the point-in-time features used in training. Metadata and pointers stored in version control or a catalog then tie the feature snapshot, dataset version, and container version to the same model version.
Model Registry on Gemini Enterprise Agent Platform is the central system for tracking and managing model artifacts, including binaries and metadata, across versions. Model Registry stores, organizes, and deploys model versions seamlessly, so multiple AutoML (opens in a new tab) versions trained from the same managed dataset can be compared and promoted without losing lineage. Training a custom model and an AutoML model from the same managed dataset lets teams compare performance before choosing which registered version moves to inference.
Model Registry depends on upstream versioned inputs to make deployments repeatable. The model version depends on the container version in Artifact Registry, the dataset version in BigQuery or Cloud Storage, and the feature version in Feature Store. When a new training run completes, its artifact is registered as a new version rather than overwriting the production version, which preserves the ability to validate performance and roll back quickly when required.
Exam tip: Use Model Registry for model binaries and model metadata, Artifact Registry for container images and pipeline components, and Feature Store for versioned feature data — they version different layers of the same release.
Automated deployment with pipelines and CI/CD uses Agent Platform Pipelines for orchestration and Cloud Build for build, test, and deploy automation. Agent Platform Pipelines automates every step from data preparation to model training, evaluation, and deployment, with scheduled pipeline runs for recurring work, execution metrics for workflow monitoring, and reusable pipeline templates for standardized workflows. Pipeline compilation is integrated as part of the CI/CD flow and execution is driven on a schedule with Cloud Scheduler, so the same path from training to inference runs consistently across environments.
Cloud Build ensures infrastructure, dependencies, and model packaging meet deployment requirements before a version reaches serving. ML releases need validation beyond application code tests:
During testing, key metrics like latency, error rate, and resource utilization are monitored through Cloud Monitoring (opens in a new tab) to identify bottlenecks. Agent Platform Pipelines handles automated model evaluation, BigQuery handles large-scale data analysis, and continuous training enables automated retraining on new data with current base models selected from Agent Platform Model Garden.
Safe and controlled model releases are deployment strategies that detect issues early, validate performance in production, and enable fast rollback to a prior Model Registry version. Automation reduces manual effort and standardizes processes, which leads to consistent deployment across environments and more reliable outcomes as systems scale. Faster iteration is balanced against release risk by validating a new version before it fully replaces the current serving version.
Model Monitoring on Gemini Enterprise Agent Platform provides post-deployment reliability by detecting data drift, tracking performance, and identifying anomalies in production. For generative AI systems, monitoring also tracks shifts in output quality and safety compliance, including evaluations of text quality, image fidelity, or code functionality. When monitoring signals degradation, the versioned lineage in Model Registry, Feature Store, and Artifact Registry lets teams trace the serving version back to its data, features, and code, then roll back or trigger retraining through the automated pipeline.
Configure and Optimize AutoML Training Pipelines is the task of defining and parameterizing AutoML jobs for data, features, objectives, tuning, cost, and security. On Gemini Enterprise Agent Platform (opens in a new tab), training is orchestrated by the AutoML TrainingPipeline resource, which parameterizes the full path from data to a production-ready model. The TrainingPipeline resource always executes the training task, and optionally exports data from a Gemini Enterprise Agent Platform Dataset, uploads the resulting Model, and evaluates the Model.
Data source configuration defines which managed data becomes the training input and how it is versioned and reused. The pipeline can export data from a Dataset resource directly into training, so the same managed dataset can train both a custom model and an AutoML model to compare performance under identical inputs. For best performance and support, the underlying data should live in Google Cloud storage systems before it is registered as a dataset.
Supported sources for training data include peer storage options with different access patterns:
Data flows from the storage service to the Dataset resource, then from the Dataset export step to the TrainingPipeline training task. To keep training repeatable, version feature data in Agent Platform Feature Store, version datasets in BigQuery or Cloud Storage, store metadata and pointers in a data catalog or version control system, and register output models in Model Registry on Gemini Enterprise Agent Platform.
Feature engineering in AutoML tabular pipelines is implemented as a sequence of pipeline components where each component runs in a container and its output artifacts become inputs to the next component. Gemini Enterprise Agent Platform Pipelines (opens in a new tab), which provides managed Kubeflow and TensorFlow Extended (TFX) services, orchestrates these stages from data preparation through training, evaluation, and deployment. This structure lets teams standardize transformations, reuse templates, and track metadata across runs.
An AutoML tabular workflow moves data through an ordered sequence:
Dataset export into statistics generation and example creation to profile schemas and detect anomalies.For forecasting, ProphetTrainerOp provides the specialized trainer, while AutoMLForecastingTrainingJobRunOp, AutoMLImageTrainingJobRunOp, and AutoMLTabularTrainingJobRunOp create the corresponding model resources. Preprocessing for these inputs can use Dataflow for Apache Beam, Managed Service for Apache Spark for Spark jobs, BigQuery for SQL transforms, and TensorFlow Data Validation (TFDV) for schema validation. For generative tasks, the same pipeline pattern adds prompt engineering, response filtering, human-in-the-loop evaluation, and data augmentation to improve prompt diversity and robustness.
Model objective selection matches task complexity and computational need to the training mode. AutoML handles simpler machine learning tasks such as tabular classification and regression, image classification, and forecasting without custom training code, while Agent Platform custom training handles custom model building for complex classification, regression, or generation work. Model Garden (opens in a new tab) provides pretrained foundation models for common text, image, and code generation tasks that can be used directly or fine-tuned for a specific use case.
| Training mode | Best for | Compute and control tradeoff |
|---|---|---|
| AutoML | Simpler tabular, image, and forecasting tasks with automated search | Low setup effort; platform controls architecture and tuning within the parameterized objective |
| Agent Platform custom training | Custom architectures, large language models (LLMs), and diffusion models | High control; configure Cloud Tensor Processing Units (TPUs) or GPU resources to accelerate training and inference at higher cost and complexity |
| Model Garden | Common applications starting from state-of-the-art foundation models | Fastest time to value; fine-tune the base model when domain performance requires it |
The training job depends on the selected mode: an AutoML job depends on the TrainingPipeline parameterization, while a custom job depends on a containerized application. A custom training job first provisions one virtual machine (VM) for single-node training or multiple VMs for distributed training, then runs the containerized application on those VMs, and only after completion deletes the VMs. For generative fine-tuning and reinforcement learning from human feedback (RLHF), labels must accurately reflect quality, relevance, and safety of generated outputs, and synthetic data from BigQuery DataFrames or platform generation can supplement insufficient or imbalanced data.
Hyperparameter tuning runs multiple trials of the training application using different hyperparameter values to discover the optimal values within a specified range. The developer defines the search space as ranges to test, and the platform searches that space, tracks each trial with autologging for experiment tracking, and selects the configuration that maximizes the objective. This separates the search definition, which the user controls, from the search execution, which the platform manages serverlessly.
In AutoML tabular pipelines, tuning is distributed across Stage1TunerOp for initial architecture and hyperparameter search and CvTrainerOp for cross-validated training of candidates. Training best practices require the training code to load input data from Google Cloud services, use platform environment variables, export model artifacts to the designated location, and remain resilient to VM restarts so interrupted trials can resume. Use the same Dataset for tuned AutoML runs and baseline custom runs so tuning gains are measured against data differences rather than data drift.
Exam tip: The AutoML TrainingPipeline always runs training, but Dataset export, Model upload, and Model evaluation are optional steps in the same resource.
Operational optimization for AutoML training pipelines combines scheduling, automation, monitoring, and enterprise security to balance speed, cost, and reliability. Gemini Enterprise Agent Platform Pipelines automates the lifecycle with scheduled pipeline runs for recurring retraining, execution metrics for monitoring workflows, reusable pipeline templates for standardized designs, Control Flows for conditional gates, pipeline compilation integrated into CI/CD, and execution via Cloud Scheduler. Training execution can optimize scheduling goodput with Dynamic Workload Scheduler or reservations, register every candidate in Model Registry, and monitor production quality with Model Monitoring on Gemini Enterprise Agent Platform (opens in a new tab) for data drift, performance degradation, and, for generative systems, shifts in output quality and safety compliance.
Enterprise security and governance bound what the pipeline can access and where data can move:
Continuous training integrates Cloud Build for building, testing, and packaging, Agent Platform Pipelines for automated evaluation, BigQuery for bulk evaluation across diverse datasets, and Dataflow validation for feature checks, with stress testing under realistic load and monitoring of latency, error rate, and resource utilization in Cloud Monitoring (opens in a new tab).
Evaluate and Interpret AutoML Model Performance is the process of reviewing the metrics and artifacts produced by automated training to decide if a model is accurate, fair, and ready for production. On Gemini Enterprise Agent Platform (opens in a new tab), the AutoML TrainingPipeline orchestrates this work by exporting data from a Vertex AI Dataset, executing training, uploading the resulting Vertex AI Model, and generating evaluation outputs. Those outputs flow into Model Registry on Gemini Enterprise Agent Platform for versioning and into Vertex ML Metadata for lineage, so every metric can be traced back to its data, parameters, and model artifact.
AutoML Evaluation Outputs and Lineage link training data, model binaries, and evaluation results into one traceable record. Data flows from the source Dataset to the training task, then to an uploaded Model and its evaluation artifact, which records how the model performed on held-out test data. Experiments on Agent Platform integrates with Vertex ML Metadata to log and build linkage across parameters, metrics, and dataset and model artifacts across many runs. This lineage lets teams compare different architectures, input datasets, hyperparameters, and hardware choices to determine what works best before promotion to Model Registry.
Aggregate metrics summarize overall model quality on test datasets, including accuracy, precision, recall, and loss tracked over time. A model is diagnosed by looking beyond a single score and evaluating performance against particular subsets of data, different model versions, and different model architectures. Bulk evaluations assess model performance across diverse datasets, while stress tests and data-integrity checks reveal failures that aggregate scores hide:
When performance drops on a specific slice, version, or load condition, the cause is typically skewed training data, mismatched evaluation data, or an architecture choice that does not generalize, which signals a need for more representative data, retuning, or continuous training on new data.
Vertex AI TensorBoard is the managed service for visualizing and comparing specific AutoML and custom experiments, including loss and accuracy over time, model graph structure, and projection of embeddings to a lower dimensional space. Experiments on Agent Platform complements TensorBoard by preserving the run-level record of which dataset version, hyperparameter set, and model artifact produced each metric. Together they depend on Vertex ML Metadata as the system of record, so a comparison in TensorBoard can be traced back to the exact inputs that created it. For repeatability, automated evaluation is run as a stage of Agent Platform Pipelines (opens in a new tab), with each pipeline component running in a container where output artifacts become inputs to subsequent evaluation components.
Vertex Explainable AI provides feature attributions that explain why a model generates a prediction by detailing the importance of each input feature used for that prediction. The attributions help validate whether the model relies on sensible signals, build trust in its behavior, and surface bias when a sensitive or spurious feature dominates decisions. Interactive analysis runs in a Vertex AI Workbench instance, which includes the What-If Tool (WIT) for interactively analyzing models for bias and the Language Interpretability Tool (LIT) for visually exploring natural language processing model behavior. Vertex Explainable AI supports custom-trained models based on tabular and image data, so explanation coverage depends on the data modality of the model under review.
Exam tip: Vertex Explainable AI feature attributions are documented for custom-trained models based on tabular and image data — do not assume the same built-in support for all AutoML video, text, or generative outputs.
A deployment-ready AutoML model is a versioned Model Registry entry whose evaluation metrics, explanation results, and infrastructure behavior all meet production constraints. Model Registry on Gemini Enterprise Agent Platform stores, organizes, and versions model binaries and metadata so candidate versions can be compared and rolled back, while Model Monitoring on Gemini Enterprise Agent Platform (opens in a new tab) detects data drift, tracks performance, and identifies anomalies in production. Serving validation also checks operational thresholds, including inference latency where AutoML text, video, or tabular models are expected to tolerate latencies greater than 100ms. After deployment, the control loop continues from Model Monitoring back to retraining through Agent Platform Pipelines for continuous training, keeping the model current as data and business needs evolve.
The AutoML TrainingPipeline resource always executes the training task, while exporting data from a Dataset, uploading the resulting Model, and evaluating the Model are optional steps in the same resource. The pipeline can export data from a Dataset directly into training, so the same managed dataset can train both a custom model and an AutoML model to compare performance under identical inputs.
Model Registry on Gemini Enterprise Agent Platform versions model binaries and metadata, Artifact Registry versions container images and pipeline components, and Agent Platform Feature Store versions feature data with BigQuery as the offline store. Together they version different layers of the same release so a deployed version can be reproduced, compared, and rolled back.
Data flows from the Dataset into the TrainingPipeline for training, from training to Model Registry for versioning, and from Model Registry to an endpoint for low-latency requests or to a batch prediction service for bulk scoring. After deployment, production traffic and predictions flow from the serving target to Model Monitoring for drift, performance, and anomaly detection.
Last updated