The growing prevalence of multimodal data streams-such as video, images, and text-poses new challenges for stream processing systems. Existing systems are designed for structured data and struggle to support multimodal workloads efficiently. While multimodal large language models (MLLMs) provide powerful semantic understanding, naively integrating them into streaming pipelines can incur excessive computation and degrade performance. This demonstration presents SAM(center dot) S & amacr;RA, a novel super-optimizer that enables real-time processing of multimodal data streams by aggressively specializing query plans to the query and data stream. Through an interactive traffic monitoring scenario, the demo demonstrates how multimodal queries are optimized by transforming naive query plans into optimized query plans. Users can directly observe how semantic, logical, and physical optimizations interact, reduce redundant model invocations, and impact system behavior, including responsiveness and execution accuracy. The demonstration shows that efficient multimodal stream processing requires rethinking query optimization beyond naive MLLM integration.
For decades, database software has reacted to hardware. AI-for-Systems is about to shorten the development time gap between them. This panel asks the DB community what genuine HW-SW co-design should look like in the coming decade, and whether we are organized to do it.
Formulating effective filtering conditions for database queries often requires a deep understanding of the specific values stored within table columns, which can impose a significant cognitive burden on users, especially when the textual columns contain hundreds or even thousands of unique values. To address this challenge, we introduce FilterPilot, an LLM-powered interactive assistant designed to adapt filtering predicates to table content. Given user-provided filtering terms and optional time constraints, FilterPilot assists by suggesting appropriate filtering conditions. At its core, FilterPilot employs a novel iterative recall-then-verify paradigm, combining LLM-based query reformulation with table-value feedback to dynamically expand search terms. This process is further optimized through an Upper Confidence Bound (UCB) selection mechanism and an early stopping strategy. We release the source code, datasets, and a live demo system on https://github.com/wusw14/FilterPilot to encourage community adoption.
The rapid evolution of AI has led to increasingly large and complex foundation models with trillions of parameters and heterogeneous architectures, making efficient and reliable training and inference systems essential. Training such models requires large-scale distributed computation, effective overlap of computation and communication, sophisticated parallelization strategies, and robust fault-tolerant mechanisms. In addition, the rise of multi-turn and autonomous agents drives new demands for asynchronous reinforcement learning and post-training systems. Inference systems, on the other hand, must support diverse workloads with varying service-level agreements (SLAs), rapidly integrate engineering optimizations, and carefully balance trade-offs among throughput, latency, cost, and availability, particularly in distributed environments. In this talk, I will discuss the major systems challenges involved in building large-scale foundation models, drawing on our experiences developing Qwen large language models and Wan video generation models. I will also present ongoing research and system co-designs for improving training, post-training, and inference at scale, with the goal of making complex AI workloads in cloud environments more efficient, agile, and reliable.
Large language models (LLMs) are increasingly integrated into relational data processing via operator-like components—entity matchers, semantic filters, data imputers, and semantic rankers—that we call LLM-Enhanced Relational Operators (LROs). A growing number of systems and components have proposed diverse LROs, yet no unified view exists. This tutorial establishes a formal taxonomy along three dimensions—operating logic, operand granularity, and implementation variant—distinguishes standalone LRO components for data preparation from multi-LRO semantic query systems, covers optimization techniques (model cascades, semantic batching, cost-based planning), examines benchmark methodology, and discusses open challenges toward deeper LLM-database co-design.
