MLOps is an extention of DevOps (development and operations) practices of putting in production machine learning (ML) models. It is focused on automation and monitoring at all the steps of ML system construction: creating reproducible pipelines, reusable software environment, testing, integration, deployment and model performance monitoring.
There are many additional components in MLOps in comparison to DevOps, due to different nature of Data Science and Software development projects. In Data Science:
- many different programming languages and frameworks are used, thus the projects don't have monolithic structure.
- there is an experimentation step during development of models, where the performance of the models and used datasets need to be tracked.
- testing needs to include the model, data and the software components.
- pipelines can be long and complex and deploying them can require automating many steps that were done manually during the construction of the system.
- once in production, the performance of the model needs to be constantly monitored, since change in incoming data can change decrease the performance. In this case, the model should be re-trained.
MLOps is a practice for collaboration and communication between data scientists and operations professionals to help mannage production ML lifecycle. Similar to the DevOps and DataOps appoaches, MLOps looks to increase automation and improve the quality of production ML while also focusiong on business and regulatory requirements.
Categories: Big Data, Containers Orchestration, Data Engineering, Data Science, Tech Radar | Tags: Data Engineering, Deep Learning, CI/CD, Data Science, Deployment, Docker, GitOps, Kubernetes, Machine Learning, MLOps, Open source, Python, TensorFlow
Apache Liminal is an open-source software which proposes a solution to deploy end-to-end Machine Learning pipelines. Indeed it permits to centralize all the steps needed to construct Machine Learning…
Mar 31, 2021
Categories: Big Data, Data Engineering, Data Science, Learning | Tags: Beam, Data Engineering, Deep Learning, Pipeline, CI/CD, Data Science, Deployment, Machine Learning, MLOps, Open source, Python, TensorFlow
Putting Machine Learning (ML) and Deep Learning (DL) models in production certainly is a difficult task. It has been recognized as more failure-prone and time consuming than the modeling itself, yet…
Mar 5, 2021
Building Machine Learning (ML) models is a time-consuming process. It requires expertise in statistics, ML algorithms, and programming. On top of that, it also requires the ability to translate a…
Dec 10, 2020
Introduction and principles of MLflow With increasingly cheaper computing power and storage and at the same time increasing data collection in all walks of life, many companies integrated Data Science…
Mar 23, 2020
“Enterprise Machine Learning requires looking at the big picture […] from a data engineering and a data platform perspective,” lectured Justin Norman during the talk on the deployment of Machine…
Sep 30, 2019