ANALYSIS OF THE APPLICATION OF KUBEFLOW AS AN AI/ML EXPERIMENTATION PLATFORM IN A CLOUD NATIVE ENVIRONMENT

• Main Author: Agustin Putri, Dwina
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/82271
• Institution: Institut Teknologi Bandung

This research aims to analyze the application of Kubeflow as an AI/ML experimentation platform in a cloud native environment, focusing on the ease of use of the interface and the multi-user feature for collaborative experiment access. The background of this research is based on the need for a platform that can facilitate AI/ML experiments efficiently and effectively in a dynamic and scalable environment. With the rapid advancement of AI/ML technology, there is an urgent need for a platform that can integrate various stages of experiments, from model development, testing, to model management. Built on Kubernetes, Kubeflow offers solutions with features such as notebooks for developing AI/ML models, Kubeflow Pipelines for tracking experiments, and integrated model management, thereby simplifying the complexities of AI/ML model development. This research uses an experimental method involving the configuration and testing of various Kubeflow features and comparing them with the MLflow platform. The primary focus of this research is to test the ease of use and effectiveness of the Kubeflow interface and evaluate the capability of the multi- user feature for collaborative experiment access. Testing was conducted through beta testing involving ten respondents. Each respondent was given a questionnaire that included questions about their experience using both platforms, particularly in terms of ease of finding features, interface intuitiveness, and accessibility from various devices. The test results show that Kubeflow has a significant advantage in the ease of finding features. Respondents stated that the Kubeflow interface is more intuitive and makes it easier for them to navigate and find the required features. The higher average score on questions related to the ease of finding features in Kubeflow compared to MLflow supports this. Additionally, the intuitiveness of the Kubeflow interface also received better ratings from the respondents. The simpler and more structured interface makes users feel more comfortable and quickly adapt to this platform. The multi-user feature in Kubeflow also received positive feedback from users. The testing involved a scenario where two users, User A and User B, collaborated on an experiment. User A created a new experiment and ran the pipeline , while v User B accessed the experiment and added comments or changes. The test results show that the multi-user feature in Kubeflow allows effective collaboration, with each user able to access and contribute to the same experiment in real-time. Furthermore, isolation and security testing showed that each user could only access and modify data within their own namespace, ensuring data privacy and security. In terms of accessibility from various devices, Kubeflow also performed well. Respondents stated that they could easily access Kubeflow from various devices, including laptops, tablets, and smartphones. This is very important in a dynamic and flexible work environment, where AI/ML researchers and developers often have to work from various locations and devices. In conclusion, this research concludes that Kubeflow better meets the criteria of a user-friendly interface and an intuitive interface compared to MLflow. Kubeflow's advantages in ease of finding features, interface intuitiveness, and accessibility from various devices make it a more user-friendly and effective platform for AI/ML experimentation in a cloud native environment. Additionally, the multi- user feature that enables real-time collaboration and ensures data security makes Kubeflow a better choice for teams working collaboratively. These findings are expected to contribute to the development of better AI/ML platforms in the future and support the wider adoption of AI/ML technology in various fields.