GPUs have evolved into powerful, cost-efficient engines for general-purpose parallel compute. This talk introduces Sirius, a GPU-native SQL engine developed in partnership between UW-Madison and NVIDIA. Sirius enables drop-in GPU acceleration for DuckDB and other SQL databases, without changing the user interface. By treating the GPU as the primary execution engine, Sirius achieves strong performance by leveraging the massive compute... Read More

What does it take to add custom data types and indexes to a database that was never designed for them? Over the past year, we built an extensibility framework for MySQL — and discovered, one by one, every place the codebase assumes it knows all the types that will ever exist. This talk walks through the design decisions that survived... Read More

We present Floe, a SQL query service designed for modern data lakehouse architectures. Floe combines three core components: FloeSQL, a compute engine for executing sophisticated SQL over open table formats; Floecat, an open-source catalog-of-catalogs that unifies metadata across Apache Iceberg and Delta Lake ecosystems; and Floescan, an indexing service that enables fine-grained data skipping within Parquet files. Floe is motivated... Read More

Multigres is an adaptation of Vitess for PostgreSQL that enables horizontal scaling and enterprise-grade operational capabilities for Postgres deployments. The system employs a layered proxy architecture with MultiGateway for query routing and MultiPooler for connection pooling, allowing organizations to manage large distributed database clusters while maintaining PostgreSQL compatibility. Key features include modeless connection pooling that works as if you're directly... Read More

This talk explores what happens when you move all the infrastructure and logic of a fully featured, real-time backend application into the database. Using SpacetimeDB and BitCraft as a case study, the talk argues that the ideas behind stored procedures deserve a modern revival as a first class database feature: bringing deployment, transactions, relational data, and general-purpose procedural logic into... Read More

Pixeltable is a novel OLTP DBMS specifically designed for multimodal AI applications. It goes beyond the traditional roles of a DBMS — data storage and retrieval — by incorporating data generation into the table schema. Pixeltable natively supports multimodal data types (e.g., video, images, audio, documents) in the context of traditional multi-user transaction semantics. Unlike analytics-focused databases, Pixeltable also allows... Read More

We spent years optimizing database internals, only to have our performance eaten up by the one thing we can’t control: the OS. General-purpose kernels are great for desktops, but for high-performance data infrastructure, they become the bottleneck. In this talk I will introduce QuinineHM, a specialized "Hardware Manager," as a replacement for traditional operating systems in database workloads. We will... Read More

Databases and other data systems have successfully democratized data-oriented computation across domains, thanks to decades of research in system internals and end-user interfaces. However, such systems center on structured (i.e., tabular) data; unstructured data—the vast majority of data—has largely been ignored. Large language models (LLMs) now give us a building block for unstructured data analysis, and we face the same... Read More

YugabyteDB is an AI-ready, multimodal, distributed PostgreSQL database. It effectively bridges the gap between traditional RDBMS reliability and NoSQL flexibility, offering a cloud-native architecture that is resilient by design. Developers can thus leverage familiar PostgreSQL semantics while achieving the elastic scale and resilience required for mission-critical applications. This talk explores YugabyteDB's unique approach to fusing PostgreSQL’s query engine with a... Read More

The shift to multi-tenant clouds and growing data demands exacerbate contention on shared data infrastructure. In these environments, contention remains a primary bottleneck to performance. While there is extensive research on concurrency control protocols, these approaches share a fundamental limitation: they handle conflicts only after they have materialized, missing opportunities to improve performance by avoiding conflicts altogether. My research addresses... Read More