Yeah, yeah: it’s officially a trend. A great buzzword for the new brand in the IT world. Quick search in Google Trends indicates that the word pulls slowly but surely. Andrew Ng himself, one of the most respected names in the world of artificial intelligence, just fell new specialization focused on it.
But … what are MLOps? Let’s make a quick overview.
I keep it simple. By quoting AWS re-invented in 2020 MLOps is:
Incorporating an experimental machine learning model into a production system.
This is the same definitionOn you can find for example On Wikipedia. The idea of delivering pieces of software to production in an automated and painless way is not new at all: DevOps is all about that. A few years ago, some people started telling the world if all companies are now, at least in part, software companies, all companies should take care of good software engineering practices.
DevOps is not just for Dev and Ops. It’s about running the whole company; It’s about the whole corporate culture of caretakers (who promise to keep the building clean) and the CEO (who promises to keep the company funded and payroll). The theory of promises has emerged as an intellectual framework to support this change in culture. And about speed, we discuss those changes.
(Mike Loukides, O’Reilly, 2014)
There are a lot of articles between DevOps and MLOps that explain (take for example This from Marco Susilo), so I’m not going to delve into it, but no doubt the same idea now prevails: more and more companies are adding machine learning or in-depth learning models to their products or using these models to help them make decisions. So, more and more companies have to make sure that he gets the expertise to perform such tasks.
MLOps has jurisdiction over three major areas:
- Information: MLOps does not have to worry about converting or cleaning data properly. it is more of a Data Scientist task. But they have to take care of how the DS gets the data, how that data is extracted, how it is processed if necessary, etc.
- Model:Once again, MLOps don’t have to worry about modeling or the differences between a random forest and a linear regressor because data scientists do it. MLO organizations should take care of how the model is sent to the cloud, how it is tested once in the cloud, what the selected metrics are, and how to record or draw them, e.g. For example, models tend to rot (This is called model decomposition), and it is critical to define a way to monitor model performance to know when it will be replaced or retrained.
- Code: This is the key. All (or most) of the work described above can now be done through the user interface. Either AWS, GCP, or Azure provide a versatile and user-friendly interface where you can move to manage your tools. However, it is difficult to maintain. And if the project grows, it just can’t be renovated. So everything has to be code. Tekee Terraform Ring the bell?
The MLOps workflow is pretty similar to any software development workflow, and yes, that’s another hint: everything is software.
- Design: The first step always. Collecting requirements, examining the problem from an infrastructure perspective, etc.
- Development: Pipeline construction, integration, close collaboration with both the DS team that builds the model and the SE team that builds the application that leverages both the model and the infrastructure. The main goal is to maximize the automation of all processes.
- Tracking: On the next day. Everything went smoothly, cheers! Now MLOps need to make sure everything is under control and that if something ceases to be under control, it is fixed or at least reported to ASAP.
I wanted to name this song MLOps commands (pretty epic, right?), But I realized that maybe the commands are too big words – I shamelessly stole it this excellent article from Manon Sikkes and I can’t compete with it. Suppose the list below is top five concerns about MLOps:
- Automation: For both ML and standard CI / CD software tubes, everything should be as automated as possible.
- Versioning: Checking changes, reverting to previous versions, etc. Applies to data, templates, and code.
- Testing: Enough said.
- Tracking: Model decomposition, architectural bottlenecks, load peaks…
- Reproducibility: Applies to the ML side. “But it worked on a laptop“It is not acceptable. Data pre-processing, feature design, model training and evaluation – everything should be reproducible.
In summary, The MLOps culture seeks to bring the DevOps and ML worlds closer together. No one knows what the secrets of the future are, but we can expect it to continue as a trend in the near future.