Pinterest, Facebook, and Twitter are very well-known ML users. They use ML to improve content delivery, form chatbot armies, and curate timelines, among many other applications. Do you think they employ MLOps, too?
What Is MLOps?
MLOps or ML Ops, short for “machine learning operations,” refers to practices to deploy and maintain ML models in production reliably and efficiently. The term is a combination of “ML” and “DevOps.”
Simply put, MLOps ensures that ML models come out of production lines as expected, work as designed, and are created in the most efficient way possible.
Here’s a video showing what MLOps is:
How Does MLOps Work?
MLOps ensures that ML models are tested and developed in separate experimental systems to ensure that they work. Before they are launched, data scientists and DevOps and ML engineers test them to ensure they are ready for production.
Like DevOps, MLOps finds ways to increase automation and improve the quality of production of ML models while adhering to business and regulatory requirements. MLOps began as a set of best practices. Now, however, it is slowly becoming a standalone approach to ML lifecycle management.
MLOps applies to an ML model’s entire lifecycle. It covers ML model generation, orchestration, deployment, and testing against health, diagnostics, governance, and business metrics.
DevOps versus MLOps, What’s the Difference?
While MLOps takes many of its cues from DevOps, the two have some notable differences. We noted down some below.
MLOps Requires More Types of Experts Than DevOps
DevOps teams are mostly made up solely of software engineers. That is not the case for MLOps teams. An MLOps team includes data scientists and ML researchers to take care of exploratory data analysis, model development, and experimentation—processes that are not tested in DevOps.
The following table shows what the different experts involved in MLOps do.
|DevOps||Building a generic application|
Standard set of libraries for specific use cases
|Unit testing||Software engineers|
|MLOps||Building a model to feed inferences|
Broad scope of tools, languages, and libraries
|Model performance (error rate)||Data scientists|
Note that in MLOps, the DevOps code, validation, and experts are necessary as well.
MLOps Requires More Tests Than DevOps
Apart from unit and integration testing in DevOps that is also applied to MLOps, testing ML systems also involves model validation, model training, and other tests.
CI/CD Testing Is Not Enough for MLOps
Continuous integration/Continuous delivery (CI/CD) testing is the process of testing software throughout its development to ensure that it works as programmed before it gets launched. In MLOps, apart from testing the ML-enabled application, every ML model built into it is tested until it works as designed.
The DevOps CI testing that involves testing and validating code and components should also include testing and validating data, data schemas, and models in MLOps.
The CD testing in DevOps that tests a single software package or service is extended to include testing an ML model that should automatically deploy another service or roll back changes. In simpler terms, every ML-enabled component of the program is tested separately, apart from the application in its entirety.
Finally, MLOps requires continuous testing (CT), which is not done in DevOps. This process involves automatically retraining and serving ML models.
The following diagrams depict what DevOps and MLOps typically involve:
Based on the diagrams alone, you can see that MLOps is more complex than DevOps. In the simplest terms, DevOps is about code. ML, meanwhile, not only deals with code but also data.
What Are the Benefits of MLOps?
MLOps provides ML model creators several benefits that include:
- Unifying the release cycle for ML and software applications
- Enabling the automated testing of ML artifacts or the outputs created by the training process, covering data validation, ML model testing, and ML model integration testing
- Allowing the application of agile principles to ML projects
- Supporting ML models and datasets separately to ensure they work, apart from the software as a whole
- Reducing the technical debt or missing components across ML models
- Ensuring a language-, framework-, platform-, and infrastructure-agnostic process
What Companies Use MLOps?
Several companies use MLOps to date. LUSH, Uber, and Netflix are some of the most popular.
- British handmade cosmetics retailer, LUSH, uses MLOps for product classification.
- Uber, which connects users to drivers and delivery riders quickly and efficiently, uses MLOps to enhance its app’s matching capability constantly.
- Streaming service provider Netflix uses MLOps to constantly improve its movie and show recommendation ability for its millions of users worldwide.
Other MLOps use cases include personalization (Holiday Extras), fraud detection (Ocado Retail and Revolut), and loan approval and credit scoring (Carbon).