03-orchestration

3. Orchestration and ML Pipelines

3.0 Introduction: ML pipelines and Mage

• What is MLOps
• Why we need to operationalize ML
• How Mage helps MLOps
• Example data pipeline

3.1 Data preparation: ETL and feature engineering

• Ingest raw data
• Prepare data for training
• Build training sets
• Data validations using built-in testing framework

3.2 Training: sklearn models and XGBoost

• Reusable training set data product
• Training pipeline for sklearn models
• Training pipeline for XGBoost
• Tracking training metrics with experiments

3.3 Observability: Monitoring and alerting

• Dashboard for sklearn training pipeline health
• Dashboard for XGBoost model explainability
• Dashboard for model training performance
• Alerts for pipeline runs

3.4 Triggering: Inference and retraining

• Automatic retraining pipeline
• No-code UI input fields to interact with models
• Inference pipeline for real-time predictions

3.5 Deploying: Running operations in production

• Setup AWS permissions and credentials
• Terraform setup
• Initial deployment to AWS
• Use GitHub Actions for CI/CD to automate deployment to production

Quickstart

1. Clone the following respository containing the complete code for this module:

   git clone https://github.com/mage-ai/mlops.git
   

2. Change directory into the cloned repo:

   cd mlops
   

3. Launch Mage and the database service (PostgreSQL):

   ./scripts/start.sh
   

4. The subproject that contains all the pipelines and code is named unit_3_observability

Notes previous editions

• 2022 Prefect notes
• 2023 Prefect notes