---

So far you’ve learned:

Day 1 → What MLOps is

Day 2 → ML lifecycle

Day 3 → Data engineering basics

Day 4 → Data drift and model decay

Day 5 → Experiment tracking and model registry

Today we answer a key question:

How does a model that was trained on a laptop actually get used inside real products like food delivery apps, banking apps, or hospital systems?

This process is called Model Deployment.

Deployment means:

Making a model available so other applications can use it and make predictions.

Without deployment, a model is just a research project.

---

1. Why Deployment Matters So Much

Think of a doctor who knows the perfect treatment but never actually treats any patient.

That knowledge is useless. Same with ML. Companies build ML models to:

• reduce fraud

• improve customer experience

• save costs

• increase revenue

• improve safety

But value only comes after deployment. This is where MLOps and DevOps meet strongly.

---

2. Two Main Ways Models Are Used in Production - Deployment patterns

---

A. Batch Inference (Predictions Done in Bulk at Intervals)

This means predictions are generated on a schedule:

• hourly

• daily

• weekly

Not instant.

Real-World Examples

Banking: Credit scores updated once per month

E-commerce: Customer churn score calculated weekly

Insurance: Risk score calculated nightly

Hospitals: Daily patient readmission risk score

Food Delivery: Daily sales demand forecasting

How It Works

Data collected

↓

Model runs on entire dataset

↓

Results stored in database

↓

Applications use stored results later

This is useful when:

• real-time is not required

• performance cost is high

• decisions are periodic

---

B. Real-Time (Online) Inference

This means the model predicts instantly when a request comes.

Real-World Examples

Banking: Fraud detection when you swipe a card

Food Delivery: ETA prediction when you place an order

E-commerce: Product recommendations while browsing

Hospitals: Emergency risk prediction during admission

How It Works

Request comes

↓

Model runs immediately

↓

Prediction returned in milliseconds

This requires:

• fast systems

• high availability

• monitoring

• scalable infrastructure

MLOps helps ensure all of this works reliably.

---

3. How Models Are Typically Exposed — Using APIs

Most real-time models are deployed as APIs (Application Programming Interfaces).

Think of an API like a restaurant counter. You place an order The kitchen prepares it

You receive the food Similarly, Application sends data, Model processes it, Prediction returned

APIs are built using tools like:

• FastAPI

• Flask

• Django

• gRPC

The API is then packaged and deployed using:

• Docker

• Kubernetes

• Serverless platforms

This is where DevOps skills blend with MLOps.

---

4. Where Models Are Deployed — Environment Choices

Models can be deployed in:

On-Premise Servers

Used by:

• banks

• government

• regulated industries

Because of data privacy.

---

Cloud Platforms

Such as:

• AWS

• Azure

• GCP

Useful for:

• scalability

• flexibility

• global users

---

Edge Devices

Such as:

• mobile phones

• IoT devices

• hospital machines

• smart cameras

Examples:

• speech recognition on phone

• face unlock

• medical device monitoring

Deployment strategy depends on business needs.

---

5. Packaging Models — Why Docker Is So Important

Docker is like a tiffin box where everything required by the model is packed:

• code

• libraries

• model file

• configurations

So it runs the same everywhere: developer laptop, staging, production

This avoids

“works on my machine” problems.

DevOps already uses Docker widely,

MLOps continues the same culture.

---

6. Deployment Pipelines — CI/CD for ML

In DevOps, CI/CD automates:

• building software

• testing

• deploying

In MLOps, CI/CD automates:

• testing ML code

• validating model accuracy

• checking data quality

• deploying model safely

Typical flow:

Code commit

↓

Build container

↓

Run tests

↓

Validate model

↓

Deploy to staging

↓

Deploy to production

This ensures safe, repeatable releases.

---

7. Deployment Strategies (Explained Simply)

Companies rarely deploy models recklessly. They use careful rollout methods, such as:

---

A. Blue-Green Deployment

Two environments exist:

• Blue → current live version

• Green → new version

Traffic switches only when tested.

If problems occur, traffic returns to Blue.

---

B. Canary Deployment

New model is tested on a small percentage of users first.

• If results are good, it is rolled out to all users.

• If not, it is rolled back.

---

C. Shadow Deployment

New model runs in parallel

but predictions are not shown to users.

Engineers compare results silently.

If stable → promote to production.

This is very common in banks and hospitals.

---

8. What Can Go Wrong During Deployment?

Many risks exist, such as:

• latency too high

• wrong predictions

• biased decisions

• accuracy drops

• data pipeline failures

• scaling issues

• compliance violations

That is why monitoring is essential after deployment. This is not optional.

---

9. Monitoring Deployed Models

Once live, models must be tracked for:

• uptime

• response time

• error rate

• accuracy

• drift

• bias

• business KPIs

Example:

Food delivery monitors:

• ETA accuracy gap

• customer complaints

• delivery delays

Bank monitors:

• fraud prevented

• false alerts

Hospital monitors:

• true detection rate

• patient outcomes

This keeps systems safe and reliable.

---

10. Rollback — Safety Net for Models

Just like software releases, sometimes models must be rolled back quickly.

Reasons include:

• incorrect results

• bias detected

• accuracy dropped

• legal risk

• data pipeline failure

MLOps ensures rollback is:

• fast

• safe

• traceable

Again, similar to DevOps best practices.

---

11. A Simple Real-World Story

A bank deploys a new fraud detection model. At first, it performs well. Then support calls increase:

“Why are my transactions getting blocked?”

Monitoring reveals:

false positives increased.

The team:

• rolls back to previous version

• investigates

• retrains correctly

Because tracking and deployment pipelines were in place,

this happened smoothly — without panic. That is MLOps maturity.

---

12. Quick Recap of Day 6

Today you learned:

• what deployment means

• batch vs real-time inference

• APIs for ML

• Docker for packaging

• cloud vs on-prem vs edge

• CI/CD for ML models

• rollout strategies

• monitoring deployed models

• rollback safety

• real-world use cases

---