Catalog
technology#Machine Learning#Platform#Analytics#Cloud

Google Vertex AI

Managed machine learning platform from Google Cloud for training, deployment and MLOps.

Google Vertex AI is a managed machine learning platform on Google Cloud for training, deploying, and operating models.
Established
High

Classification

  • High
  • Technical
  • Technical
  • Intermediate

Technical context

Google Cloud StorageBigQuery for feature engineering and data analysisCloud Monitoring and Logging

Principles & goals

Prefer managed services to reduce operational complexity.Separate training, validation and serving workloads.Versioning and traceability of models and data.
Build
Team, Domain

Use cases & scenarios

Compromises

  • Vendor lock-in through use of specific services and APIs.
  • Security and compliance risks when storing sensitive data in the cloud.
  • Model quality degradation without adequate monitoring.
  • Use automated pipelines for training, validation, and deployment.
  • Systematically version models and data.
  • Set up monitoring and alerting for model quality.

I/O & resources

  • Training data (e.g. CSV, TFRecord) in Cloud Storage
  • Model and training configuration
  • Service accounts and permissions
  • Versioned model in the model registry
  • Deployed endpoints with metrics
  • Monitoring and drift reports

Description

Google Vertex AI is a managed machine learning platform on Google Cloud for training, deploying, and operating models. It provides scalable training jobs, integrated MLOps tooling, model monitoring, and online/batch inference. As a cloud service, Vertex AI reduces infrastructure overhead and enables fast iteration and integration with Google Cloud services.

✔

  • Reduced infrastructure management via managed services.
  • Scalable training and serving options for production workloads.
  • Integrated MLOps tooling for pipelines and monitoring.

✖

  • Dependency on Google Cloud and proprietary services.
  • Costs can grow quickly for large-scale training.
  • Limited control over underlying infrastructure details.

  • Training time

    Time consumed to complete a training run.

  • Prediction latency

    Average response time of an online endpoint.

  • Cost per training/month

    Monetary expenses for training runs and infrastructure.

Image classification for product recognition

Retail company trains a model for automatic product classification and uses Vertex AI for training and serving.

Customer churn prediction

Data science team automates feature engineering and CI/CD for re-training in Vertex AI.

NLP model for support tickets

Support process integrates a text classification model with online inference via Vertex AI endpoints.

1

Set up Google Cloud project and enable billing.

2

Configure service accounts and roles for Vertex AI.

3

Place data in Cloud Storage, define and run a training job.

⚠️ Technical debt & bottlenecks

  • Insufficient automation for re-training and tests.
  • Non-versioned feature pipelines.
  • Manual steps in deployment and rollback.
Data I/OCost managementModel lifecycle management
  • Storing sensitive personal data without encryption.
  • Promoting models trained on unbalanced or unverified labels to production.
  • Ignoring drift alerts and failing to respond.
  • Hidden costs from long training runs and GPU usage.
  • Complexity due to region-specific services and data locality.
  • Lack of governance for model access and permissions.
Knowledge in machine learning and model validationExperience with Google Cloud Platform and IAMAbility to integrate CI/CD for MLOps
Scalability for large training and serving workloadsSecurity and access control for model assetsAutomatability of MLOps processes
  • • Requires a Google Cloud project and billing enabled.
  • • Some services are region-dependent.
  • • Access and compliance requirements for data storage.