101 of MLflow

Here’s everything you need to know about the open source MLOps framework in 60 secs:

Origins

MLflow was created by Matei Zaharia, the now CTO of databricks. As such, MLflow is primarily supported by databricks. Databricks only supports Mlflow when it comes to MLOps (the operationalizing of machine learning models by machine learning engineers.)

MLflow was launched in 2018 to address the complexities of the whole development and production process of machine learning models. The idea is to provide one framework to rule all steps.

Ownership

Since 2020, MLflow has been under the stewardship of the Linux Foundation.

The Key Components of MLflow

Going from 0 to production with a machine learning task means lots of different stages, and MLflow has different components to help you manage all of them.

1. You got data? Great. If you train your first model, you’ll want logging. If you run your models through the mlflow library, you get autologging.

2. You got a model and want to evaluate it? Great, The MLflow tracking server gives you automatic evaluation of it.

3. You got a second model? Great! First, the Tracking server allows you to compare models to find the best one, and the Model Registry allows you to register and retrieve different versions of it!

4. Finally, to get your model into production, MLflow helps you to create a REST endpoint (e.g. using Flask) and package everything into a container you can then send onto kubernetes or databricks.