MLOps is an acronym for Machine Learning Operations and consists of a compilation of best practices in Software Engineering in order to put ML models into production. This is a very short description of what MLOps is, but for now it is just enough.
The term “Production” is referred to the stage of a ML model lifecycle where it becomes available for the end user and every interaction that this user makes with the system triggers the model calling and it returns a response (through an API, for example)
In order to demonstrate the usage of MLOps platforms, this course will be based on the development of a model which is aimed to predict the time of a taxi ride in NY using data available on the NYC Taxi Data repository.
As mentioned before, MLOps is a set of best practices to put ML in production. And these best practices are related to the ability of removing human interaction more and more by creating what we call pipelines.
It is very common during experimentation stage to use Jupyter notebooks, however these sources of code can become quite complex to understand and eventually, create problems if we override something that was not supposed to.
So, creating pipelines allows us defining specific tasks to be run in order and generate a final result without surprises.
Generally, a pipeline consists of the following macro stages:
Once the model is deployed, our system becomes available for usage.
Nowadays, there are plenty of frameworks that can be used, like MLFlow, Prefect, Kubeflow and this course will cover all their main functionalities!
We can define 5 levels of model maturity: