Abstract: In this talk, I will explore a new direction for data management: the synthesis of workload-specific database systems, which I call Bespoke DBMSs. Rather than designing a single general-purpose engine to serve a wide range of workloads—for example, one engine for all OLAP workloads or one for all OLTP workloads—I argue that the code-generation capabilities of large language models (LLMs) and coding agents make it possible to automatically construct database engines tailored to individual workloads, such as TPC-C or TPC-H. While LLMs can already generate individual code fragments, synthesizing an entire database system presents a fundamentally different systems challenge: how can AI agents be guided from generating individual building blocks—such as storage formats, indexes, operators, and execution logic—to assembling, testing, and validating them as a complete, coherent, and executable system? I will present my vision for Bespoke DBMSs and a concrete end-to-end synthesis pipeline that takes a workload specification and generates Bespoke OLAP engines. Through initial experimental results, I will show that synthesized database systems can significantly outperform general-purpose engines on their target workloads, illustrating the potential of moving from one-size-fits-all to one-size-fits-one database systems. I will discuss the opportunities and challenges of this new paradigm and, more broadly, what it means for systems research when database architecture itself can be generated, specialized, and continuously reinvented by AI.

Speaker Bio: Carsten Binnig is a Full Professor in the Department of Computer Science, where he leads the Data and AI Systems group. His research lies at the intersection of databases and AI, focusing on leveraging AI models and hardware to build the next generation of data systems. His work has pioneered areas including learned cost models, semantic data systems, and database systems designed for AI hardware. He received his Ph.D. from the University of Heidelberg in 2008 and subsequently worked as a postdoctoral researcher at ETH Zurich and in industry, where he contributed to the development of in-memory database systems. He has also been a visiting researcher with the Google Systems Research Group. Professor Binnig is a founding member of hessian.AI and leads the Systemic AI for Decision Support research area at the German Research Center for Artificial Intelligence (DFKI) in Darmstadt. His research has received multiple Best Paper and Best Demo awards at leading database conferences, including ACM SIGMOD, VLDB, and CIDR, and he was awarded a LOEWE Top Professorship by the State of Hesse.

Abstract: Cloud VM allocation policies are at the core of a hyperscale fleet, yet they are still largely handcrafted by expert engineers through years of incremental tuning. While deep learning approaches such as reinforcement learning offer a path toward automated optimization, their black-box behavior and difficulty of diagnosing and controlling failures make it challenging for adoption. This talk explores whether agentic systems can provide a more interpretable and practical path toward automated policy optimization. Using cloud VM allocation as a case study, I will discuss how agents can combine system knowledge, simulation, experimentation, and iterative discovery to explore policy spaces while making their reasoning and trade-offs visible to engineers. I will also discuss the practical challenges and open questions that remain, and what they suggest for the broader use of ML in systems engineering.

Speaker Bio: Haoran Qiu is a Senior Research Engineer at Microsoft Azure Research – Systems (AzRS). His recent work focuses on ML for systems and self-evolving agent systems, including reliable adaptive resource management, large generative model inference, and agentic cloud operations. He obtained his PhD in Computer Science from the University of Illinois Urbana-Champaign, advised by Prof. Ravishankar Iyer. He was named an ML and Systems Rising Star (2023) and is a recipient of the Mavis Future Faculty Fellowship (2022). During his PhD, he was a visiting researcher at IBM Research and worked with Google and Microsoft Research.