Vertex AI Mlops

Vertex AI MLOps

When to Use

• Designing ML training or serving infrastructure on GCP
• Setting up model monitoring or retraining pipelines
• Choosing between AutoML and custom training
• Preparing for GCP Professional Data Engineer or ML Engineer exam

Core Jobs

1. AutoML vs Custom Training

2. Vertex AI Pipelines

• Orchestrates ML workflows as DAGs (Kubeflow Pipelines or TFX)
• Each step = a containerized component (preprocessing, training, evaluation, deployment)
• Use kfp.v2 SDK or pre-built Google Cloud Pipeline Components
• Store pipeline artifacts in Cloud Storage; metadata in Vertex ML Metadata

3. Feature Store

• Centralized repository for ML features (avoid feature duplication across teams)
• Online store — low-latency serving (< 10ms) for real-time inference
• Offline store — batch access for training (BigQuery-backed)
• Features defined once, reused across models

4. Model Serving

• Endpoint — deploys one or more model versions, handles prediction requests
• Batch prediction — asynchronous, for large offline prediction jobs
• Online prediction — synchronous, for real-time serving
• Traffic splitting between model versions for A/B testing or canary releases

5. Model Monitoring

• Skew detection — training vs serving data distribution drift
• Drift detection — serving data distribution changes over time
• Alert thresholds configurable per feature
• Monitored logs sent to BigQuery for analysis

6. Model Registry

• Version all trained models centrally
• Stage models through: Experiment → Staging → Production
• Alias support for promoting/rolling back versions

Key Concepts

• ML Metadata — tracks lineage: which dataset trained which model, which pipeline produced what artifact
• Explainable AI — feature attributions (SHAP values) for model transparency
• Vertex AI Workbench — managed JupyterLab for experimentation
• Training pipeline vs custom job — pipeline = orchestrated multi-step; custom job = single training run

Checklist

• Training data versioned and reproducible?
• Model evaluation metrics gated before promotion?
• Serving endpoint has traffic splitting for safe rollout?
• Model monitoring enabled (skew + drift detection)?
• Feature Store used to avoid feature duplication?
• Pipeline steps containerized and versioned?

Output Format

• 🔴 Critical — no model monitoring in production (silent degradation)
• 🟡 Warning — no traffic splitting for new model versions, no feature versioning
• 🟢 Suggestion — Feature Store for cross-team feature reuse, Explainable AI for compliance

Exam Tips

• Feature Store online = real-time serving (low latency); offline = batch training (BigQuery)
• Model monitoring = skew (train vs serve) + drift (serve distribution over time)
• Vertex AI Pipelines = Kubeflow Pipelines on GCP (not Cloud Composer/Airflow)
• AutoML Tabular = good baseline; custom training when you need specific architecture
• Batch prediction = no endpoint needed; just submit job → results to GCS/BigQuery
• Traffic splitting on endpoints = canary release for models (same as canary deployments)