Skip to content
Deadline to Die

Phase 1: Foundations

Phase 1: MLOps Foundations

In this phase, we move from “Notebook-based Research” to “Production-grade Engineering.”

🟢 Level 1: The ML Lifecycle (CRISP-ML)

Traditional Software (DevOps) vs. Machine Learning (MLOps):

  • DevOps: Code + Configuration = Binary.
  • MLOps: Code + Data + Configuration = Model.

1. The Stages

  1. Business Understanding: Define the KPI (e.g., reduce churn by 5%).
  2. Data Acquisition: ETL from data lakes.
  3. Modeling: Experimentation and Validation.
  4. Deployment: Packaging and Serving.
  5. Monitoring: Feedback loops.

🟡 Level 2: The Reproducibility Crisis

If you can’t recreate your model from 6 months ago using the same code and data, your system has failed.

2. The 3 Pillars of Reproducibility

  • Code: Git (version control).
  • Environment: Docker/Conda (dependency isolation).
  • Data: DVC/S3 (data versioning).

🔴 Level 3: Standardizing the Workspace

3. Move away from Notebooks

Notebooks (.ipynb) are great for research but terrible for production.

  • Problem: Hidden state (out-of-order execution), hard to test, hard to version.
  • Solution: Modularize code into .py files. Use notebooks only for visualization.

4. Environment Isolation

Use Docker to ensure the model runs exactly the same on your Mac as it does on a Linux server in the cloud.