Introducing full automation and best practices to MLOps

Fully Automated MLOps

Arturo Opsetmoen Amador

Senior Consultant - Machine Learning

The MLOps lifecycle

Image of the MLOps lifecycle. Three circle cycles connected to each other. The names of each circle from left to right are: Design, Development, Deployment.

Fully Automated MLOps

Maturity levels in MLOps

$$

The maturity levels include:

  • Manual ML workflow

  • Semi-automated ML workflow

  • Fully automated ML workflow

Fully Automated MLOps

Manual ML workflow - Ad hoc experimentation

An image illustrating how models are manually delivered from development to production using a manual process.

Fully Automated MLOps

Semi-automated ML workflow

Image of data being fed to an orchestrated experiment pipeline.

Fully Automated MLOps

Semi-automated ML workflow

The previous image is extended by including code that pushed to a source repository and a feature store that feeds data to the system.

Fully Automated MLOps

Semi-automated ML workflow

The previous image is extended by including a deployment component. This deploys automated ML pipelines in the architecture. In addition a ML metadata store receives metadata generated by the automated pipeline.

Fully Automated MLOps

Semi-automated ML workflow

The previous architecture is extended. Now models produced by the automated pipeline are pushed to a model register which in turns delivers models to a model serving module. The model serving module is connected to a prediction service in the architecture.

Fully Automated MLOps

Semi-automated ML workflow

The prediction service module in the architecture is connected to a performance monitoring component which in turns is connected to an automated trigger. The automated trigger can start the automated pipeline in the system.

Fully Automated MLOps

Fully automated ML workflow

An image showing a high level overview of the fully automated MLOps reference architecture.

Fully Automated MLOps

Automation in the ML life cycle - Design

A table illustrating that activities in the design and planning phase are not automatable.

Good practices:

  • Apply reproducible processes
  • Write detailed documentation
Fully Automated MLOps

Automation in the ML life cycle - Development

A table illustrating that activities in the development phase are not all automatable.

Good practices:

  • Remember we are developing software
  • Using version control
  • Use orchestration tools
Fully Automated MLOps

Automation in the ML life cycle - Operations

A table illustrating that activities in the operations phase are mostly automatable.

Use:

  • Automated testing
  • CI/CD/CT/CM
Fully Automated MLOps

Let's practice!

Fully Automated MLOps

Preparing Video For Download...