MLOps vs DevOps

If a Notebook Session is your workbench and a Job is your factory line, then MLOps (Machine Learning Operations) is the entire supply chain management system that keeps the factory running efficiently, safely, and predictably.

In the current 2026 landscape, the MLOps lifecycle has evolved from simple model deployment into a continuous, circular system of System Orchestration.

Before we dive in, yes, I used Nano Banana to generate this beautiful diagram.

The 5 Core Stages of the MLOps Lifecycle

1. Scoping & Data Engineering

Before code is written, the business problem is defined and the data foundation is built.

• Problem Framing: Defining success metrics (e.g., accuracy, latency, or ROI).
• Data Pipelines: Automating the ingestion and cleaning of data.
• Feature Store: A central repository where "features" (processed data inputs) are stored and versioned so they can be reused across different models.

2. Experimentation (The "Sandbox")

This is where Data Scientists work in Notebook Sessions.

• Algorithm Selection: Testing different models (Random Forest vs. Neural Networks).
• Hyperparameter Tuning: Finding the best settings for the model.
• Experiment Tracking: Using tools like MLflow to log every run, so you know exactly which version of code and data produced which result.

3. CI/CD for ML (The "Pipeline")

Unlike traditional software, you aren't just testing code; you are testing Code + Data.

• Continuous Integration (CI): Testing the training scripts and validating data schemas to ensure no "garbage" enters the system.
• Continuous Deployment (CD): Automatically packaging the model into a Model Artifact (the ZIP file we discussed) and deploying it to a production endpoint.

4. Deployment & Serving

The model is now "live" and handling real-world requests.

• Model Serving: Exposing the model via a REST API.
• Inference Strategies: * Shadow Deployment: Running the new model in parallel with the old one to see how it performs without affecting users.

5. Monitoring & Retraining (The "Feedback Loop")

This is the most critical part of MLOps. Models, unlike software, "rot" over time.

• Data Drift: When the real-world data changes (e.g., a fraud detection model fails because scammers changed their tactics).
• Concept Drift: When the relationship between the data and the prediction changes.
• Automated Retraining: When performance drops below a certain threshold, the MLOps pipeline automatically triggers a new Job Run to retrain the model on fresh data.