Productionizing machine learning pipelines
Challenge
Differences between two platforms.
Under the guidance of Giuseppe D’Alessio, ING’s IT Chapter is charged with deploying machine learning pipelines into a Hadoop production platform. However, its data scientists develop their machine learning models on Hathi, an experimentation platform. Because of the differences between the two platforms, transferring code as a package between them is often difficult.
Solution
Closing the gap between ING’s data scientists and its platform team.
So ING asked Xccelerated to help the organization deploy machine learning products using Kubernetes. After a full month of data engineering boot camp, Xccelerated’s young professionals (who already had several years of software development experience under their belts) were ready to join ING’s existing Hathi platform team. Since then, Xccelerated’s data engineers have helped close the gap between ING’s data scientists and its platform team. First, we created awareness about using containers in machine learning environments by planning and facilitating a training session focused on Kubernetes.
By sharing knowledge about Docker (including how to write, build and run Docker images) we were able to go in depth about how to create Kubernetes pods to gain insights on deployments and services.
Result
Production-ready projects in an experimentation platform.
Within the first quarter, a successful proof of concept for deploying machine learning models with Kubernetes was created. Using Python, the team developed a tool that continuously integrated with GitlabCI by default, so the data scientists could create production-ready projects in an experimentation platform while constantly testing the code. And since the tool was developed in Python, everyone can contribute new features. Users can also access their logs from GitlabUI, monitor their deployments and easily detect any issues. Manual steps are automated in the process.
Our next step? We’ll be teaching more data scientists and platform engineers on how to use the new tools. Meanwhile, the happy collaboration between ING and our data engineers has led to permanent offers.
Other references
