MLOps explained

The importance of MLOps and the processes associated with it

Credit: Microsoft Azure

Outlines

1. What is MLOps
2. Why do we need MLOps
3. What are the benefits of MLOps

1. What is MLOps

• Model development
• Continuous integration + deployment (CI/CD)
• Monitoring
• Validation
• Governance

MLOps Goals

1. Faster experimentation and model development
2. Faster deployment of updated models into production
3. Quality Assurance

2. Why do we need MLOps?

• Experimentation
• Track metrics
• Source control in the code
• Checkpoint steps in the ML lifecycle
• Automating proper validation / staged deployment
• Monitoring model performance efficacy
• Automated retraining

ML pipeline for data scientist teams

• Data
• Train
• Validate
• Deploy
• Monitor

Version the source data and its attributes

1. Traditionally take up the most amount of time to clean up and get in shape.
2. Data can come in different formats and different sources.
3. The better the quality of data, the better the quality and efficacy of the model.

Build the model

• Feature selection / generation
• Algorithm selection
• Hyperparameter tuning
• Fitting the model
• etc.

Learn from mistakes

The data is just as crucial and informs the next set of combinations to try.

Responsible ML

• Understand: Interpretability fairness
• Protect: Differential privacy confidential machine learning
• Control: Aduit trail datasheets

Model Drift

Model was defined by the business case needs, if there is a new business case needs, it will need a rethink or retraining of the model.

Data Drift

When the model is trained on the demographics of a set of users and the population it is being utilized on doesn’t match that same demographic.

• Seasonality
• Consumer preferences
• New products

Automatically triggers model retraining, so that the new model can cater to the new requirements.

3. Quality Assurance

• Create reproducible ML pipelines
• Enable reusable ML environments
• Register package and deploy models
• Capture governance data
• Generate alerts
• Monitoring ML applications
• Automation

What MLOps can provide?

• Scalability and management
• Reusability and reproducibility
• Effortless CI/CD
• Maintain model health
• Advocates responsible AI practices

Best practices

1. Create models with reusable ML pipelines
2. Automation is key for robust MLOps
3. Monitor performance
4. Monitor data drift and utilize the insights to retrain the model
5. Enable automatic audit trail creation for all artifacts