We are excited to announce the fourth night of talks in the NYC Systems series in 2026! Talks are agnostic of language, framework, operating system, etc. And they are focused on engineering challenges, not product pitches.
We are pleased to have Manya Bansal and Sugu Sougoumarane speak, with a guest lightning talk from Zhenghong Yu, and glad to have Trail of Bits as a partner for the venue.
Manya Bansal is a third-year Ph.D. student at MIT. Her research lies at the intersection of programming languages, compilers, and high-performance computing, with a focus on designing programming systems that make GPU programming modular, composable, and efficient. She received Distinguished Paper Awards at PLDI in 2023 and 2026 and was awarded the NVIDIA Graduate Fellowship. She is currently a Member of Technical Staff at humans&, where she has successfully traded Boston winters for San Francisco fog.
To achieve peak performance on modern GPUs, one must balance two frames of mind: issuing instructions to individual threads to control their behavior, while simultaneously tracking the convergence of many threads acting in concert to perform collective operations like Tensor Core instructions. The tension between these two mindsets makes modular programming error prone. Functions that encapsulate collective operations, despite being called per-thread, must be executed cooperatively by groups of threads.
In this talk, I will present Prism, a new GPU programming language designed to restore modularity without sacrificing low-level control over collective operations. Prism’s central abstraction is typed perspectives, which encode at the type level the granularity at which a programmer is reasoning about and controlling thread behavior. By making these perspectives explicit, Prism ensures that operations are only executed with an appropriate view of the underlying hardware resources, unlocking compositional GPU programming without compromising performance.
Sugu Sougoumarane is Head of Multigres at Supabase and the co-creator of Vitess, the database clustering system he originally built at YouTube to scale MySQL and which now runs behind some of the largest deployments on the internet. He later co-founded PlanetScale around it. He returned from a three-year sabbatical to start Multigres.
Multigres is an adaptation of Vitess for PostgreSQL, built as a layered proxy that sits in front of the database: MultiGateway for query routing, MultiPooler for connection pooling, and a high availability system that honors the rules of consensus protocols rather than approximating them. He has written an extended series on generalized consensus and on a more flexible Paxos.
Postgres has become the default choice for a very large number of new systems, and the ceiling everyone eventually hits is the same one: you run out of machine. The options after that are read replicas, application-level sharding, or moving off Postgres, and none of them are good.
This talk covers Multigres, which takes the architecture Vitess proved out on MySQL and rebuilds it for the Postgres ecosystem. I will go through the layered proxy design, modeless connection pooling that behaves as though you are connected directly to Postgres, and a relational sharding model that co-locates related data so that entire joins can be pushed into a single shard.
Compatibility is the constraint that shapes everything, because it is the thing the Postgres community values most and the thing sharding most wants to take away. I will be honest about which tradeoffs we accepted and which ones we are still arguing about.
Zhenghong Yu is a PhD student in the Department of Computer Sciences at the University of Wisconsin–Madison, advised by Prof. Paris Koutris. His research focuses on database theory and on applying theoretical results to practical system design, alongside a broader interest in operating systems and distributed systems. He received his B.E. from ShanghaiTech University. His recent work includes FlowLog, an efficient and extensible Datalog engine for static analysis.
Datalog is a natural fit for static analysis: write down the recursive rules and let the engine figure out how to run them. The problem is that existing systems often force a choice between performance and extensibility—highly optimized engines are difficult to modify, while more flexible designs can come with substantial overhead.
This lightning talk covers FlowLog, a Datalog compiler designed to get both. FlowLog compiles Datalog through an optimized relational representation into dataflow, providing fast and scalable execution while making it possible to extend the same system with incremental computation, performance profiling, and semantics beyond standard Datalog.