MLOps Lifecycle & CRISP-ML Framework

🎯 Learning Objectives

• Understand the CRISP-ML framework for ML projects
• Identify the stages of the ML lifecycle
• Distinguish between MLOps maturity levels
• Recognize key stakeholders and their roles

📋 Prerequisites

• Basic understanding of ML model development
• Software engineering concepts

---

1. 📖 Core Content

1.1 What is MLOps?

1. Standardize: Create reproducible ML pipelines
2. Automate: Reduce manual handoffs between teams
3. Monitor: Track model performance in production
4. Govern: Manage model versions, access, and compliance

1.2 CRISP-ML Framework

1.3 MLOps Maturity Levels

1.4 Key Stakeholders

---

📝 Practice Questions

Q1

<strong>Q1</strong>: What's the difference between traditional DevOps and MLOps?

MLOps extends DevOps to handle the unique challenges of ML: data dependencies, model evaluation, and ongoing monitoring. Q2

<strong>Q2</strong>: At MLOps maturity level 0, a data scientist trains a model and sends the file to an engineer for deployment. List three problems this causes.

1. Reproducibility: No record of training data version, hyperparameters, or environment → hard to debug or retrain
2. Manual handoff: Emailing files, forgetting dependencies, dependency hell
3. No monitoring: Model behavior in production is unknown → silent degradation
4. Version conflicts: Multiple models, no tracking of which is deployed where
5. Rollback difficulty: Can't easily revert to previous model version

These problems are why organizations invest in MLOps infrastructure. Q3

<strong>Q3

<strong>Q3</strong>: A model that predicts customer churn has 85% accuracy in testing but 72% in production. What could go wrong?

Possible causes:

1. Data drift: Customer behavior changed (seasonal, economic, etc.)
2. Training-serving skew: Different preprocessing in training vs production
3. Concept drift: The relationship between features and churn changed
4. Sampling bias: Training data wasn't representative of production data
5. Feature staleness: Some features aren't available in real-time

This is why monitoring is essential — without it, you wouldn't detect the 13% accuracy drop. Q4

<strong>Q4

<strong>Q4

<strong>Q4</strong>: At minimum, what should a team at maturity level 1 implement?

At level 1 (DevOps but no MLOps):

1. Git: Version control for all code
2. CI/CD pipeline: Automated testing and deployment of code changes
3. Environment standardization: Docker or conda for reproducible environments
4. Basic artifact storage: Save model files with timestamps

Even without full MLOps automation, these practices ensure code quality and deployment reliability. The next step is adding data and model versioning.

</details> * * * ## 🔗 Cross-References - **Next**: [Experiment Tracking](/notes/04-degree-electives-bsda5014-mlops-week02-02-experiment-tracking) - **Video**: BSDA5014 Week 1 transcripts [Join Discord](https://discord.gg/gE2m4Qrdqv) [Next**Experiment Tracking**](/notes/04-degree-electives-bsda5014-mlops-week02-02-experiment-tracking)