MLOps Guide by Aguimar Neto (MLOps) is an ML framework to maintain, deploy, and monitor the end-to-end life cycle of ML applications . Similar to DevOps, it aims at shorter code-build-deploy loops, high-quality, reproducibility, and automated deployment with monitoring. However, MLOps is distinct in several ways: it requires different tools and is more complex to implement than DevOps .
It also introduces several processes that can be difficult to automate such as model retraining, data validation, and model management. To address these challenges, MLOps introduces a new pipeline component called Continuous Training. This pipeline automatically retrains the model to adapt to new incoming data. To initiate this process, triggers are used. MLOps also provides features such as data and model archiving and management, which make it possible to track the history of changes and revert to previous versions of the application.
Streamlining Machine Learning Operations: Aguimar Neto’s Expert Insights
In this case study, we implement an MLOps level 2 pipeline for a time-series forecasting application in the hourly day-ahead electricity market using a transformer-based ML model. This MLOps system ingests the relevant data from Fingrid and the Finnish Meteorological Institute to raw data storage, selects essential features from these data, validates them for missing values, transforms data to a format needed by the prediction service, and stores the transformed data in a feature store for reusability. In addition, this MLOps system is integrated with a web interface to explore predicted values.