Chapter 34: Introduction to MLOps
MODULE 6 | MLOps
Chapter 34: Introduction to MLOps
This chapter is the foundation of the whole module. We will see what MLOps is , why a notebook model does not work in production , what the whole life (lifecycle) of an ML project looks like, and which problem each tool in this module solves. At the end we will learn how to save and load a model, because that is the first step of deployment.
TOPIC CHECKLIST (tick them off as you go)
-
What MLOps is and why it matters
-
☐
-
☐ The difference between a notebook and production
-
☐ The ML lifecycle: data, training, deployment, monitoring
-
☐ Map of this module's tools (roadmap)
-
☐ Saving and loading a model with
joblib -
☐ Setting up the project folder and virtual environment
34.1 What is MLOps and why does it matter?
MLOps (Machine Learning Operations) is the name for the methods and tools by which an ML model is built, run and maintained in a reliable way . It is a blend of three worlds: Machine Learning (building the model), DevOps (running software without problems) and Data Engineering (keeping data clean and ready).
34.2 The Difference Between a Notebook and Production
| Thing | Notebook (your computer) | Production (the real world) |
|---|---|---|
| Who uses it | Only you | Thousands of people, apps, other systems |
| How the model runs | Run a cell, see the answer | Through an API,24 hours a day, automatically |
| Environment | Your library versions | Must bethe sameeverywhere (Docker) |
| When something goes wrong | Run the cell again | Check the logs, get an alert, roll back |
| When the model gets outdated | You never find out | You find out throughmonitoring |
| Record of experiments | Memory or notebook comments | Parameters and metrics inMLflow |
The most famous sentence: "It was working on my computer." The goal of MLOps is to put an end to this sentence.
34.3 The ML Lifecycle
An ML project is never built once and finished . It keeps running in a cycle . Here are its stages:
DIAGRAM
DATA -> TRAINING -> EVALUATION -> DEPLOYMENT -> MONITORING
DATA -> TRAINING -> EVALUATION -> DEPLOYMENT -> MONITORING
^ |
|________________ new data, new model ___________________|
| Stage | What happens | Tool in this module |
|---|---|---|
| Data | Collect, clean and version the data | Modules 1 to 4 (pandas, sklearn) |
| Training | Train the model, keep a record of experiments | MLflow(Chapter 37) |
| Evaluation | Check accuracy on test data, compare models | MLflow UI, pytest (Chapters 37, 38) |
| Deployment | Run the model as an API, pack it in a box | FastAPI,Docker(Chapters 35, 36) |
| Automation | Automatic tests and build on every change | GitHub Actions(Chapter 38) |
| Monitoring | Health of the live model, logs, data drift | logging, drift check (Chapter 39) |
34.4 The First Step: Saving and Loading a Model
Before running a model in an API or Docker, you have to save it to a file . Otherwise training would have to be repeated every time. For scikit-learn models, the most common way is joblib :
import joblib
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier
X, y = load_iris(return_X_y=True)
model = RandomForestClassifier(n_estimators=50, random_state=42)
model.fit(X, y)
joblib.dump(model, "iris_model.joblib")
print("Model saved")
loaded = joblib.load("iris_model.joblib")
sample = [[5.1, 3.5, 1.4, 0.2]]
print("Prediction:", loaded.predict(sample)[0])
OUTPUT
Model saved
Prediction: 0
Prediction 0 means setosa (the first species of Iris). Note this: the model file is trained only once , and after that the API or Docker only loads it.
34.5 Project Setup: Folder and Virtual Environment
Make a separate virtual environment for every project so that libraries do not clash with each other. Do this in the terminal:
TERMINAL
mkdir mlops-practice
cd mlops-practice
python -m venv venv
# Windows:
venv\Scripts\activate
# Mac / Linux:
source venv/bin/activate
pip install fastapi uvicorn scikit-learn joblib pandas httpx pytest mlflow
pip freeze > requirements.txt
VERSION NOTE
This module uses Python 3.12 . Some libraries (such as TensorFlow ) do not yet support very new Python versions (for example 3.14), and newer library versions can bring small changes. If any code gives an error, first check python --version and pip list. In this module, the code was tested with FastAPI 0.159, scikit-learn 1.8, MLflow 3.17, pandas 3.0 .
34.6 Common Mistakes and Hands-on
Hands-on
-
Save the model of one of your old projects with
joblib, and load it in a new notebook and make one prediction. -
Make a
mlops-practicefolder and a virtual environment, and runpip freeze > requirements.txt. -
For one of your old projects, write: which stage of the lifecycle has it reached (data, training, deployment, monitoring)?
-
Write in one sentence: why is "It works on my computer" not enough in production?
KEY TAKEAWAYS
-
MLOps = ML + DevOps + Data Engineering: building, running and maintaining a model in a reliable way .
-
In a notebook only you run the model; production needs an API, 24 hours a day, and the same environment .
-
The ML lifecycle is a cycle : Data, Training, Evaluation, Deployment, Monitoring, then back to Data.
-
joblib.dump/joblib.loadsave and load the model; load only trusted files . -
Write versions in
requirements.txtso that the same environment is built everywhere.