This repo will provide insights on the DevOps workflow for Machine Learning. This will content will help partenrs deliver art of the possible conversations with their customers. Additionally, we will have training content that can be leveraged for MLOps-in-a-Day sessions. This site is an aggregator of content and will help our partners go to one location to find all the relevant and up-to-date content.
Machine Learning has typically been an ad-hoc process where business executives request analysis and data scientists conduct it. These processes are quite brittle and create large scale technical debt. Data scientists typically ran their workshops as independent contractors who didn't need to collaborate or scale. The motto it works on my computer was sufficient since these experiments where for ad-hoc research and not end-user applications.
In today's world, we need to build out mission critical applications that contain machine learning and artificial intelligence models. These applications drive the wider adoption of ML & AI that requires enterprises to build out a DevOps practice (COE) as it relates to Machine Learning. The key drivers are;
- Multiple Data Scientists working on one or more models
- Application Developers need to reuse ML models for applications
- Reproducibility of their models
- Monitoring and Deployment of their models
- Agile releases to minimize bugs and impact on end-users
Without the DevOps practices for Machine Learning, customers can be negatively impacted by these models and tarnish a company's brand image.
Microsoft product engineering team has published a Github project that contains a collection of MLOps examples and solutions. This should be your main landing page for the latest technical assets related to MLOps.
Jordan Edwards, Program Mananger for MLOps, was a guest on TWIML where he talked about Enterprise Readiness, MLOps and Lifecycle Managment.
TWIMLAI = This Week in Machine Learning and AI
Microsoft Mechanics
- https://github.com/chronicle17/DSDevOps
- https://docs.microsoft.com/en-us/azure/architecture/reference-architectures/ai/mlops-python (Reference Architecture)
- https://rsethur.github.io/MLOps/ (MLOps Recipes)
- https://github.com/Azure-Samples/MLOpsDatabricks (Databricks)
There is a PowerPoint presentation authored by Mark Tabladillo and Jason Virtue. The slide deck supports introductory conversations ("art of the possible") for new end-users. Go to the slides folder for the pptx.
Here is a sample that will support a hands-on showcase of Azrue MLOps. The setup instructions and URL for the Azure DevOps Demo Generator
Microsoft Cloud Workshops are guided learning materials that can be conducted by partners to their end-customers. These sessions have a whiteboard session to design overall architecture and a guided lab to build out the technical asset.
- MLOps Architecture Deep Dive