Until a few years ago, to leverage data, companies hired only Data Scientists and Machine Learning Engineers. These professionals could build predictive models, allowing companies to leverage locally and make (some) important decisions.
However, Machine Learning projects failed when they were supposed to go into production. Companies missed opportunities and customers were dissatisfied.
Data Scientists focus exclusively on building Machine Learning Models. Once in the hands of the end user, there is no system in place to ensure that these models will work properly in the real world and in an environment different from the one in which they were trained.
However, the real world is unpredictable and constantly changing. Therefore, the performance of a Machine Learning model can change drastically from one day to the next.
For example, the slightest change in the training database can affect the accuracy of the model. This phenomenon is called “data drift” and must be detected quickly to update the model before it becomes biased.
Similarly, seasonality causes regular and predictable changes to the data at specific time intervals. Machine Learning Models need to be regularly updated with these seasonal changes.
In addition, many Machine Learning models are not suitable for production use because they cannot handle the large volumes of data entering the system in real time.
These phenomena, related to a lack of procedures for using machine learning models, can have an extremely negative impact on the performance of algorithms in production. To address this problem, the profession of MLOps (Machine Learning Operations) was created.
An MLOps has both Machine Learning and operations skills. His job is to support the workflow that follows the construction of the Machine Learning Model.