Stop 17 · MLOps & Model DeploymentTopics · 5/10

CI/CD for ML Pipelines

In short

In normal software, CI checks the code. In ML, a bad dataset or a worse model can slip through with perfect code. ML CI/CD adds checks for data and model quality too.

Automating testing, validation and deployment for ML code and models — extends standard CI/CD with ML-specific checks like data validation and model performance gates.

The blob in a lab coat holds up a milk bottle to inspect it beside a conveyor belt of bottles.

Think of it like this

The dairy quality lab

Before milk packets leave a dairy, the lab tests every batch for fat and freshness. A bad batch never reaches the shop, even if the packing machine works perfectly. ML CI/CD is that lab: code tests, data checks and an accuracy gate before any model ships.

Watch it happen

From git push to a deployed model

Push

Unit tests

Data check

Train+eval

Gate

Deploy

1/6You push a change to the feature code. GitHub Actions starts the workflow.

The big ideas

What you need to know

Test code as usual

Feature functions and API handlers still need normal unit tests.

Validate the data

Check schema, ranges and null rates on every run. Bad data is the most common silent ML failure.

Gate on model quality

Fail the pipeline if the new model is worse than a threshold or than production.

Retrain on schedule

Many teams also run the pipeline on a timer or when new data lands, not only on code pushes.

See it in code

Walk through the code

yaml

1name: ml-ci2on: [push]3jobs:4  test-and-gate: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 validate_data.py data/train.csv14      - run: python train.py --out model.joblib15      - run: python evaluate.py model.joblib --min-accuracy 0.85

1/4Run on every push to any branch.

A GitHub Actions workflow. Each run step must exit 0, so evaluate.py just exits with an error if accuracy is too low.

Save yourself the pain

Common mistakes

Oops

Gating only on 'training finished without errors'.

Do this instead

Training can finish and still produce a worse model. Gate on a real metric against a fixed validation set.

Oops

Committing a 500 MB dataset into the git repo for CI.

Do this instead

Store data in object storage or a data versioning tool like DVC, and pull it in the workflow.

Your mission

Add a quality gate

~40 min

Quick check

Your code tests pass, but last week's data export had all labels set to 0. Which check catches it?

Go deeper

1Set up a CI pipeline that runs tests and a basic model-performance check on every push.
2Learn what should gate a deployment: code tests plus a minimum accuracy threshold.
3Automate retraining or redeployment triggers where it makes sense for your project.