Learn to test code, data and models in CI, gate promotion on validation metrics, and retrain safely with GitHub Actions, DVC and MLflow.
It is a Friday afternoon at acme-shop. A data scientist retrains the acme-fraud model by hand on her laptop, sees a better precision number in her...
A pipeline is only as strong as its cheapest check. Run the fast checks first so a typo fails in 20 seconds, not after a 12 minute training run.
A gate is a script that exits with a number. Zero lets the pipeline continue. Anything else stops it.
The workflow glues everything together. We keep it small, pinned and unprivileged, and we split it by trust.
Passing the gate on a pull request means "this change looks safe to merge".
Three things must happen to version N before a human is asked to approve it: an honest test number, a scanned image and a recorded digest.
It is a Friday afternoon at acme-shop. A data scientist retrains the acme-fraud model by hand on her laptop, sees a...
A pipeline is only as strong as its cheapest check.
A gate is a script that exits with a number. Zero lets the pipeline continue. Anything else stops it.
The workflow glues everything together. We keep it small, pinned and unprivileged, and we split it by trust.
Passing the gate on a pull request means "this change looks safe to merge".
Three things must happen to version N before a human is asked to approve it: an honest test number, a scanned image and...
By now version N has a gate report, a frozen test evaluation and a scanned image.
ML pipelines have the usual CI risks plus three of their own: sensitive training data, powerful credentials and...
Passing the gate means the model looked good on past data.
Continuous training means the pipeline starts itself.
You will add the trusted evaluator and gate to the shared mlops-lab project, prove it works, break it four ways on...
Quick reference Common mistakes The first mistake is letting the candidate score itself.
Roles that use the skills in this module.
AI Engineer
MLOps Engineer
Platform Engineer
DevOps Engineer
Not a software engineer sheet. Every problem comes from real DevOps, SRE, Platform and Cloud interviews, from your first script to a system you build yourself.
Open the Coding SheetCI/CD ships code changes: it tests them and releases them. Continuous training (CT) ships a new model when the data or the world changes, even if no code changed. An ML system needs both, because a model can go stale while the code stays perfect.
No. A gate runs many times, and every run that reads the test set nudges you to tune against it. Use the validation set for gates and baselines, and evaluate the test set once per release candidate. Record that fact somewhere atomic and durable, and enforce separate release thresholds on the result.
A gate is only as honest as the code that produces the scores it reads. If a pull request can change the scoring script, it can write any scores it likes. Run the evaluator, the split logic and the gate from main, and let the candidate control only the model file.
Only when the gate is strong, the blast radius is small and rollback is one command. For high-impact models such as fraud, most teams keep a human approval step on the final alias move. Automate the checks and the evidence, then let a person press the button.