Vertex AI Best Practices Advanced

This lesson covers production-grade best practices for Vertex AI, including cost optimization, security, monitoring, scaling, and organizational strategies for managing ML workloads at scale.

Cost Optimization

• Use preemptible VMs for training jobs that can tolerate interruptions — up to 80% cost savings
• Right-size machine types — monitor GPU/CPU utilization and adjust accordingly
• Auto-scale endpoints with min_replica_count=0 for low-traffic endpoints
• Use batch prediction instead of online prediction for large-volume, non-real-time workloads
• Set training budgets with budget_milli_node_hours for AutoML to cap costs
• Clean up resources — undeploy unused models, delete old endpoints, and stop Workbench instances

Security Best Practices

Model Monitoring

Vertex AI Model Monitoring detects data drift and prediction anomalies in production:

# Enable model monitoring on an endpoint
from google.cloud.aiplatform import model_monitoring

objective_config = model_monitoring.ObjectiveConfig(
    training_dataset=model_monitoring.RandomSampleConfig(sample_rate=0.8),
    training_prediction_skew_detection_config=
        model_monitoring.SkewDetectionConfig(
            data_source="bq://project.dataset.training_table",
            skew_thresholds={"feature_1": 0.3, "feature_2": 0.3}
        ),
    prediction_drift_detection_config=
        model_monitoring.DriftDetectionConfig(
            drift_thresholds={"feature_1": 0.3, "feature_2": 0.3}
        )
)

# Create monitoring job
monitoring_job = aiplatform.ModelDeploymentMonitoringJob.create(
    display_name="my-monitoring-job",
    endpoint=endpoint,
    objective_configs=objective_config,
    logging_sampling_strategy=model_monitoring.RandomSampleConfig(sample_rate=0.8),
    schedule_config=model_monitoring.ScheduleConfig(monitor_interval=3600)
)

Experiment Tracking

Use Vertex AI Experiments to track and compare training runs:

• Log metrics, parameters, and artifacts for every training run
• Compare experiments side-by-side in the Cloud Console
• Integrate with Vertex AI TensorBoard for visualization
• Use experiment lineage to trace models back to their training data and code

Organizational Best Practices

• Use labels on all resources for cost allocation and resource management
• Separate environments — use different projects for dev, staging, and production
• Version everything — training code, data, models, and pipeline definitions
• Automate with CI/CD — trigger pipeline runs from code commits using Cloud Build
• Document model cards — maintain documentation for each production model including purpose, limitations, and performance metrics

Scaling Strategies