Evaluating the Effectiveness of Machine Learning in Storage Systems: Are We Taking the Right Direction?

The rapid advancements in technology have led to a significant increase in the amount of data being generated and stored. This has created a greater demand for storage systems that are efficient and reliable, especially in critical sectors like finance and healthcare. However, traditional storage solutions are struggling to keep up with the growing data volumes and the need for real-time data processing. To tackle these challenges, machine learning techniques have emerged as a promising approach to improve storage systems. Storage vendors have widely adopted these techniques to enhance their offerings. However, as we evaluate the various ways machine learning can be applied to storage, we've noticed a gap in how effectively it aligns with business objectives. In this presentation, we aim to address this gap by suggesting additional approaches that can complement existing solutions in a non-disruptive manner. Our talk focuses on three key areas. First, we delve into the philosophical aspect of machine learning interpretability. We ponder whether it is truly necessary to have a complete understanding of how black box models make decisions. Second, we explore the importance of counterfactual reasoning and what-if scenarios, particularly in risk-sensitive systems like storage security. We discuss how leveraging causal inference can provide a more comprehensive and informed perspective, leading to better decision-making processes. Finally, we introduce the concept of algorithm-agnostic uncertainty quantification using the conformal prediction framework. This framework acts as a wrapper around any machine learning model and can quantify the reliability of individual predictions, a feature that is currently lacking in most models. Throughout the presentation, we showcase the effectiveness of these focus areas through real-life storage use cases such as data security, intelligent tiering, anomaly detection, and predictive maintenance. Our goal is to empower the audience to apply machine learning techniques in their own industry settings. We provide practical insights and guidance that will enable industry professionals to navigate the ever-changing landscape of storage systems, contributing to a better and more progressive storage community