Stop 12 · Model Deployment & MLOpsTopics · 5/8

CI/CD for ML Pipelines (Overview)

In short

Retraining and deploying by hand is slow and error-prone. CI/CD runs tests, training checks and deployment automatically every time someone pushes code, so bad models get caught early.

Automating the testing and deployment of model updates — similar to software CI/CD, but also needs to validate model performance, not just that the code runs without errors.

The mascot watches a conveyor belt of milk packets where a mechanical arm pushes a leaking one aside.

Think of it like this

The dairy packing line

At a dairy plant, every milk packet rolls down a belt past checks: weight, seal, leaks. A leaking packet is pushed off the belt before it reaches a truck. CI/CD is that belt for your code and models, with automatic checks at each station.

Watch it happen

What runs on every push

Push

Unit tests

Train

Quality gate

Deploy

1/5You push code or a config change to GitHub. The workflow starts automatically.

The big ideas

What you need to know

CI = test every change

Continuous integration runs checks on each push so problems show up within minutes.

CD = release automatically

Continuous delivery ships passing builds, often after a manual approval step for production.

ML adds data checks

Unlike normal apps, ML pipelines also test data quality and model metrics, not just code.

Scheduled retraining

Pipelines can also run on a schedule, like weekly, to retrain on fresh data.

See it in code

Walk through the code

yaml

1name: ml-ci2on: [push]3jobs:4  test-and-train:5    runs-on: ubuntu-latest6    steps:7      - uses: actions/checkout@v48      - uses: actions/setup-python@v59        with:10          python-version: "3.12"11      - run: pip install -r requirements.txt12      - run: pytest tests/13      - run: python train.py --out model.joblib14      - run: python evaluate.py --model model.joblib --min-f1 0.80

1/4Run this workflow on every push.

A GitHub Actions workflow in .github/workflows/ml-ci.yml. evaluate.py should exit with an error if F1 falls below the threshold.

Save yourself the pain

Common mistakes

Oops

Retraining on the full dataset inside every CI run.

Do this instead

Use a small sample for fast checks on each push. Run full training on a schedule or on demand.

Oops

Deploying any model that trains without errors.

Do this instead

Add a quality gate: the new model must meet a metric threshold before release.

Your mission

Your first ML workflow

~40 min

Quick check

What does an ML pipeline check that a normal web app's CI usually doesn't?

Go deeper

1Read an overview of how ML pipelines differ from standard software CI/CD.
2Learn what an automated 'model performance gate' before deployment might check.
3Understand this is often owned jointly with ML/platform engineers, not solely by a data scientist.