Model Deployment & MLOps

Building a machine learning model is only 50% of the job.
The real value of ML comes when a model is deployed, monitored, updated, and scaled in production.

This is where Model Deployment and MLOps (Machine Learning Operations) come into play.

MLOps combines:

• Machine Learning
• Software Engineering
• DevOps practices

Its goal is to reliably deliver ML models into production and keep them performing well over time.

---

1. Model Serialization

What Is Model Serialization?

Model serialization is the process of saving a trained machine learning model to disk so it can be:

• Reused later
• Shared
• Deployed into production

A trained model exists in memory during training. Serialization converts it into a portable file format.

---

Why Model Serialization Is Important

• Avoid retraining every time
• Enables deployment
• Ensures reproducibility
• Saves time and compute

---

Common Serialization Formats

Pickle

• Python’s native serialization format
• Saves model objects directly

Joblib

• Optimized for large NumPy arrays
• Faster than pickle for ML models

ONNX

• Framework-agnostic format
• Used for cross-platform deployment

---

Real-World Example

A recommendation model trained on millions of user records is saved once and reused for inference.

---

2. Flask / FastAPI

Why Web Frameworks Are Needed

ML models cannot be used directly by applications.
They need a web interface so other systems can interact with them.

Flask and FastAPI help expose ML models as web services.

---

Flask

Flask is a lightweight Python web framework.

Advantages:

• Simple
• Easy to learn
• Good for small services

Limitations:

• Slower than FastAPI
• Less suitable for high-traffic systems

---

FastAPI

FastAPI is a modern, high-performance framework built on top of Starlette.

Advantages:

• Very fast
• Automatic API documentation
• Built-in data validation

FastAPI is widely used in production ML systems.

---

Typical ML API Flow

1. Receive input data
2. Preprocess input
3. Load serialized model
4. Predict output
5. Return result as JSON

---

3. REST APIs

What Is a REST API?

REST (Representational State Transfer) is an architectural style for communication between systems.

REST APIs allow clients to:

• Send requests
• Receive responses over HTTP

---

Why REST APIs Are Used in ML

• Platform independent
• Easy to integrate
• Scalable
• Language-agnostic

---

Common HTTP Methods

• GET → Fetch data
• POST → Send data for prediction
• PUT → Update model/version
• DELETE → Remove resources

---

ML REST API Example

A resume screening model receives candidate details and returns a match score.

---

4. Docker Basics

What Is Docker?

Docker is a containerization platform that packages an application with:

• Code
• Dependencies
• Runtime environment

This ensures consistency across environments.

---

Why Docker Is Important for ML

• Eliminates “works on my machine” issues
• Simplifies deployment
• Improves scalability
• Enables microservices architecture

---

Docker Concepts

Image

A blueprint containing application code and dependencies.

Container

A running instance of an image.

Dockerfile

A script that defines how to build an image.

---

ML Use Case

A Docker image includes:

• Python
• ML model
• API code
• Required libraries

---

5. CI/CD for ML

What Is CI/CD?

CI/CD stands for:

• Continuous Integration
• Continuous Deployment

It automates the process of:

• Testing
• Building
• Deploying ML pipelines

---

Why CI/CD Is Critical for ML

• Frequent model updates
• Version control
• Reduced manual errors
• Faster experimentation

---

ML CI/CD Pipeline Steps

1. Code commit
2. Automated tests
3. Model training
4. Model validation
5. Deployment
6. Monitoring

---

Popular CI/CD Tools

• GitHub Actions
• Jenkins
• GitLab CI/CD
• Azure DevOps

---

6. Model Monitoring

Why Model Monitoring Is Necessary

Once deployed, model performance degrades over time due to:

• Data drift
• Concept drift
• Changing user behavior

Monitoring ensures the model remains reliable.