What is MLOps and Why Should I Care?
Over the last 10 years, DevOps (a portmanteau of ‘development’ and ‘operations’) has arisen as a common set of practices that combines software development and IT operations. Its goal is to shorten the development lifecycle and provide faster delivery at a higher quality. The title of ‘DevOps’ however in my opinion gives a false sense of coherence to something that in reality is the accumulated body of best practices developed from thousands of lessons learned (often the hard way).
As more and more organizations take machine learning-based solutions to market, their experiences and the lessons drawn from them have spawned a similar body of knowledge. As can be seen in the trend of Google searches over the last 3 years, this knowledge has matured to the point of earning a title – MLOps.
MLOps refers to the set of practices that aims to build, deploy, and maintain machine learning (ML) models. While it resembles DevOps in many ways, the development and deployment of a ML system has several important differences from the DevOps processes. These differences include additional development artifacts, stakeholders, and requirements.
The goal of the rest of this blog is to outline what I believe are the two most important differences between MLOps and DevOps to give a general awareness of potential pitfalls and peculiarities in the process. These differences are:
- ML Code and data are independent
- ML Code is a small piece of the bigger picture
ML Code and Data are Independent
In ML systems, the original input data is often transformed and put through a feature engineering process. This is then used to train a production model. While the traditional DevOps process may produce artifacts such as the code itself and the computing environment it is intended to run in, the ML process includes additional entities that must be tracked and maintained.
In addition to the code used to prepare the data and train the model for example, the ML process introduces:
ML Code is a Small Piece of the Bigger Picture
When an ML model is in production, it is a small part of a much larger solution. If the interaction between the different components is not considered, it can result in the accumulation of hidden technical debt. Some examples of hidden sources of technical debt are:
MLOps is a large and constantly evolving body of knowledge, and its practices are already supported by commercial ML and cloud computing products. Here at Softcrylic, we have extensive experience helping our clients develop and deploy ML solutions according to best practices, including areas such as process automation, containerization, pipeline orchestration and monitoring.
If you have any questions regarding MLOps for your business, please reach out!
