Machine Learning Pipelines Made Easy for Quality Assurance Professionals

A very simple guide to how machine learning works

Machine learning looks hard. But it is not.

If you know QA, you already know the basics.

ML systems have three parts. We call them FTI:

F = Feature (clean the data)

T = Training (teach the model)

I = Inference (use the model)

Let me explain each one.

Part 1: Feature Pipeline

What does it do?

It cleans dirty data.

Simple example:

You have messy data. Names are written in different ways. Dates are in wrong formats. Numbers have errors.

This pipeline fixes all that. It makes data clean and ready.

Feature Pipeline Detail

In QA words:

You never test with bad data. You clean it first. This pipeline does the same thing.

The clean data goes to a Feature Store.

Part 2: Training Pipeline

What does it do?

It teaches the model.

Simple example:

You show the model 1000 pictures of cats. You tell it “this is a cat” each time. The model learns what a cat looks like.

In QA words:

You learn from requirements. Then you write test cases. The model learns from data. Then it can make predictions.

Picture:

The smart model goes to a Model Registry.

Training Pipeline Detail

Part 3: Inference Pipeline

What does it do?

It uses the model to answer questions.

Simple example:

Someone shows a new picture. The model says “this is a cat” or “this is not a cat.”

In QA words:

This is like running tests in production. The model is working and giving answers.

Inference Pipeline Detail

Two Important Storage Places

Feature Store

Keeps clean data

Saves old versions

Everyone uses same data

Model Registry

Keeps trained models

Saves old versions

You know which model is in production

The Full Picture

Full FTI Pipeline Overview

Why This is Easy for QA

You already know:

✓ How to check data quality → Test Feature Pipeline

✓ How to compare old vs new → Test Training Pipeline

✓ How to test in production → Test Inference Pipeline

Five Things to Remember

Three parts. Feature, Training, Inference. That’s it.

Clean data is key. Bad data = bad model.

Save everything. Keep old data. Keep old models. You can go back if needed.

Test each part. Don’t test everything together. Test one part at a time.

Your skills work here. QA testing skills work for ML testing too.

Last Words

ML is just software with a learning step.

You already know how to test software. Now you can test ML too.

Start simple. Ask: “Show me the three pipelines.”

Then test each one.

You can do this.

---