NebulaStream – Data Stream Processing for the Sensor-Edge-Cloud-Continuum

Speaker: Volker Markl (Technische Universität Berlin)
Keynote Session: Tuesday, May 5 | 8:30 - 9:30 AM
Room: Place du Canada

Abstract

Modern data-driven applications arising in such domains as smart manufacturing, healthcare, and the Internet of Things, pose new challenges to data processing systems. Traditional stream processing systems, such as Flink, Spark, and Kafka Streams are ill-suited to cope with the massive scale of distribution, the heterogeneous computing landscape, and requirements, such as timely processing and actuation. Classical approaches like managed runtimes, interpretation-based query processing, and the optimization of single queries that neglect interactions, greatly limit throughput, latency, energy-efficiency, and the general usability of these systems for emerging applications involving distributed data processing at scale in a sensor-edge-cloud-environment.

To overcome these limitations, we are researching and building NebulaStream, a novel open-source data stream processing system for massively distributed, heterogeneous environments. NebulaStream supports (potentially resource-constrained) heterogeneous devices, a hierarchical topology (with the distribution of computation and data flow in a cloud-edge-continuum), and the sharing of computations and data across multiple concurrent queries. This presentation discusses the design goals and core concepts of NebulaStream and and looks back at inspirations drawn from our prior work on Stratosphere and Apache Flink, among others.

---

Bio

Model Lake Management

Speaker: Renée J. Miller (University of Waterloo)
Keynote Session: Wednesday, May 6 | 8:30 - 9:30 AM
Room: Place du Canada

Abstract

The concept of data lakes emerged in the early 2010s to mean collections of raw, unstructured data —- as organizations recognized the untapped value of this messy data. The study of data lakes has evolved and matured shaping how we store, manage, and extract insights from these massive heterogeneous information stores. We are now seeing the emergence of model lakes, repositories of large sets of pre-trained AI models. In this talk, I argue that model lakes offer a new transformative paradigm for organizing and understanding the growing ecosystem of AI models. I review our vision for model lakes and what the application of principled data management can do for understanding and using AI models.

---

Bio

Data-Centric AI: Powering the Machine Learning Lifecycle from Preparation to Reliable Inference

Speaker: Lei Chen (Hong Kong University of Science and Technology)
Keynote Session: Thursday, May 7 | 8:30 - 9:30 AM
Room: Place du Canada

Abstract

While the recent success of Artificial Intelligence is often attributed to the "trinity" of hardware, algorithms, and data, research has historically prioritized model design over data management. However, as AI transitions into critical real-world applications, the limitations of a model-centric approach have become evident. In this talk, I will present a shift toward data-powered learning, where strategic data handling—from preparation to deployment—serves as the primary driver for system performance. We explore how sophisticated data preparation, such as self-supervised learning for spatial interpolation, can overcome data scarcity, while coreset selection and data replay strategies enable efficient stream learning and prevent catastrophic forgetting in unsupervised continual learning.

Furthermore, we examine the pivotal role of data in optimizing the efficiency and reliability of Large Language Model (LLM) inference. By implementing context-aware semantic caching and adaptive request scheduling on hybrid caches, we can significantly reduce latency and scale effective throughput. To ensure reliability, we discuss how Knowledge-Enhanced Retrieval-Augmented Generation (RAG)—utilizing advanced graph-based retrieval and skyline optimization—mitigates hallucinations and provides explicit reasoning paths. By treating data as a first-class citizen throughout the entire pipeline, we can build AI systems that are not only more efficient and scalable but also more trustworthy and precise.

---

Bio

The End of Data System Engineering As We Know It

Speaker: Tim Kraska (Massachusetts Institute of Technology)
Keynote Session: Friday, May 8 | 8:30 - 9:30 AM
Room: Place du Canada

Abstract

Machine learning (ML) and generative AI (GenAI) are fundamentally changing the way we build, operate, and use data systems. For example, ML-enhanced algorithms—such as learned scheduling techniques and optimized indexing and storage layouts—are now being deployed in commercial data services. GenAI-powered code assistants help developers build features more quickly, ML-based techniques simplify operations by automatically tuning system parameters, and GenAI assistants aid in debugging operational issues.

In this talk, I will first highlight how we transitioned several ML-for-systems techniques developed at MIT into production within Amazon. I will then discuss how generative AI is transforming the way people interact with data, highlighting selected projects from my team at AWS and MIT. Finally, I will present a new project at MIT called G5, which rethinks how we design the next generation of data systems with AI at the core. G5 introduces an AI-first development paradigm in which natural language, structured as a system ontology, becomes the primary source of truth, and code is treated as a generated artifact. This shift moves development from writing and maintaining code to specifying intent and continuously governing system behavior at a higher level of abstraction.

---

Bio