The ML Lifecycle (CRISP-ML)

🔄 The ML Lifecycle (CRISP-ML)

Traditional software development (DevOps) focuses on code. MLOps focuses on the interaction between Code, Data, and Models. CRISP-ML(Q) is the industry standard for this lifecycle.

🏗️ 1. Business & Data Understanding

Before writing a single line of Python, you must define the success criteria.

Define KPIs: (e.g., “Reduce false positives in fraud detection by 10%”).
Data Audit: Is the data available? Is it labeled? What is its frequency?

🛠️ 2. Data Preparation (The DE Step)

This is where 80% of the work happens.

ETL/ELT: Extracting from lakes, transforming with Spark/DuckDB.
Feature Engineering: Creating input variables ( $X$ ) that have predictive power.
Data Validation: Checking for nulls, types, and schema violations.

🧪 3. Modeling & Experimentation

The “Science” phase.

Algorithm Selection: XGBoost, Random Forest, or Neural Networks?
Hyperparameter Tuning: Searching for the optimal $\eta, \lambda$ , etc.
Cross-Validation: Ensuring the model generalizes to unseen data.

🚀 4. Deployment (The “Ops” Step)

Moving from a .pkl file on a laptop to a production service.

Packaging: Dockerizing the inference code.
Serving: Exposing the model via FastAPI or BentoML.
Infrastructure: Provisioning CPUs/GPUs on Kubernetes.

📊 5. Monitoring & Maintenance

The phase where most models fail.

Performance Decay: Accuracy drops over time as the world changes.
Retraining Loops: Triggering Phase 3 again with fresh data.