AI & ML

MLOps Templates

A release path for machine learning models, with approval gates and drift alerts built in.

Download the PDF

Get the MLOps Templates overview

A short PDF to share with your team. It covers what is in the package, how it works, and a worked example.

The problem

Good models stall on the way out of the notebook.

A model tests well. Then nobody can say which version is live, whether the input data has shifted, or how to roll back. Retraining happens when someone remembers.

What's included

The same release path for every model.

01

Training and release gates

Train from a config file. Each run is checked against release thresholds, gaps between groups, and the model already in production. Deploy as a batch job or behind an endpoint.

02

Registry and rollback

Versions move from candidate to production with an approver who neither trained nor registered the model. One command restores the last archived version. The registry can be swapped for MLflow or Azure ML.

03

Feature definitions

Windowed features defined once, relative to an as-of time, so training sets are point-in-time correct. A check catches data leakage.

04

Monitoring and Evaluation Cards

Population stability index against the training profile, plus missing values, new categories, freshness, accuracy, and cost. Every run writes an Evaluation Card for the AI inventory.

How it works

The release path.

  1. Train

    Fit the pipeline and write a run folder with the model, metrics by group, a data profile, the config, and the code version.

  2. Gate

    Every threshold met, gaps between groups inside their limits, and no meaningful drop compared with the production model.

  3. Register

    The run becomes the next candidate version, with its gate result attached.

  4. Promote

    Production needs a passing gate and an approver who is neither the trainer nor the registrant.

  5. Roll back

    One command brings back the last archived version. Every change is logged.

  6. Monitor

    Scheduled checks compare live data with training data and alert when a limit is crossed.

Worked example

A churn model, and the month after a price change.

A churn model is trained, gated, and promoted by a separate approver. Then the data shifts under it.

What happens
  • The trainer tries to approve their own model and is turned down
  • Customer spend rises
  • Support tickets double
  • A plan code the model has never seen shows up
Result

Monitoring flags all three shifts and exits with an alert, so a scheduled job can open a ticket or start retraining.

Where it fits

What it runs on, and where it stops.

Works with

  • Azure Machine Learning, including private online endpoints with Entra ID and blue/green releases
  • Databricks Feature Engineering and MLflow
  • GitHub Actions to train and gate on pull requests, approve for production, and check drift weekly

Assumptions and limits

  • Training templates cover binary classification. The gate, registry, and monitoring work for any model
  • The folder registry suits one team. Shared use moves to a platform registry
  • Drift means the data changed. Whether that hurts the model is the owner’s call, so alerts are reviewed together
Ask Dex

Ask about the MLOps Templates

Dex answers from our own pages and links to where it found the answer.

Have a model stuck in a notebook?

Show us how it runs today. We will map the route to production, rollback included.