During the 2021 Dataiku Product Days, Director of Sales Engineering Grant Case showed us how Dataiku can help in the deployment of MLOps. Before we dive into the demo, however, let’s go through a quick refresher on MLOps.
What Is MLOps?
To discuss any concept, it's always good to start with a definition, and MLOps is no exception. In Introducing MLOps, an O’Reilly guidebook written by the Dataiku team, MLOps is defined as:
[MLOps is] the standardization and streamlining of machine learning lifecycle management.”
How Is MLOps Different From DevOps, ModelOps, and DataOps?
- DevOps: MLOps, as its name implies, is about managing machine learning. While DevOps and MLOps share a number of characteristics and a similar name, MLOps is not DevOps. DevOps focuses on code development and deployment. MLOps includes these aspects, but also includes others such as data and governance.
- ModelOps: While you may hear them both used interchangeably today, MLOps is not ModelOps. MLOps is a subset of ModelOps. ModelOps can include other types of models, such as rule based.
- DataOps: MLOps is not DataOps. Again, MLOps is highly dependent on data. But DataOps focuses on all aspects of data which is beyond MLOps’ scope.
Now for the Serious Stuff
We went through a quick introduction of MLOps, but how do we turn these nice thoughts into reality? This demo demonstrates how Dataiku meets the needs of the final two legs of MLOps: operate and monitor. Now let's jump into Dataiku and bring some of these MLOps components to life.
MLOps and Closing the Project Loop
When developing an ML pipeline, you likely are taking an iterative process. Data is connected to, explored, and prepped, models are trained and assessed, and this process continues to some point where you believe the model is ready to be sent to deployment. For many organizations, it can take months just to get to this point. Indeed, it is a critical point, but the moment the model has finished training and deployed, it begins to decay. How do we ensure models continue performing at their best? Enter: MLOps.
MLOps closes that loop and makes ML processes simpler, more scalable, more dependable, and less risky. By incorporating all aspects of the lifecycle from plan, development, build, test, release, deploy, operate, and monitor, MLOps ensures that organizations can better leverage their existing machine learning assets and more easily create and manage new ones in a simple and more controlled manner.
Ultimately, MLOps is about making machine learning scale inside organizations by incorporating techniques and technologies, such as DevOps, and expanding them to include machine learning, data security and government. MLOps turbocharges the ability of organizations to go farther, faster with machine learning and leapfrog their competitors.