Georgia Institute of Technology
Designing ML-centric Data Systems for Efficiency and Usability
Abstract
dc:description.abstractOver the past six decades, relational databases have been remarkably successful in managing structured data. However, the growing demand for analytics over unstructured data, such as videos, images, and text, driven by modern Machine Learning (ML) workloads, exposes fundamental limitations in traditional database systems. Bridging this gap requires a new class of data systems that treat ML models as first-class citizens, integrating them directly into the query engine and providing optimizations tailored for their unique characteristics. This dissertation presents the design, implementation, and evaluation of techniques that form the foundation of ML-centric data management systems. It introduces four systems—EVA, Seiden, Aero, and PRISM—that collectively address challenges of efficiency and usability across multimodal workloads. EVA accelerates exploratory video analytics by automatically materializing and reusing the results of expensive User-Defined Functions (UDFs) through a symbolic reuse framework. Seiden revisits the long-standing “proxy model” assumption in visual databases and demonstrates that indexing directly with oracle models and exploration-exploitation sampling delivers superior execution performance and query accuracy compared to state-of-the-art video database management systems. Aero extends Adaptive Query Processing (AQP) to ML workloads by using runtime feedback to reorder predicates and scale resources dynamically, achieving better performance improvement over static optimizers. With the advent of Large Language Models (LLMs), the interface to data systems is shifting from SQL toward natural language. While the community has largely focused on accuracy for Natural Language to SQL (NL2SQL), the dollar cost of these pipelines is often ignored. To this end, PRISM is an optimization framework that treats cost as a first-class citizen and systematically navigates the cost-accuracy trade-off tailored to each specific schema and user requirement. Together, these contributions lay the foundation for the next generation of data systems designed for AI-driven workloads.
Degree
thesis:*- Name thesis:degree_name
- Computer Science, PhD
- Department thesis:degree_department
- Computer Science
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kakkar, Gaurav Tarlok
- Advisor dc:contributor.advisor
-
- Arulraj, Joy
- Committee members dc:contributor.committeemember
-
- Fatma Ozcan
- Shamkant Navathe
- Rong, Kexin
- Mussmann, Steve
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/80243