Abstract

Federated learning (FL) is a distributed machine learning paradigm that enables data owners to collaborate on training models while preserving data privacy. As FL effectively leverages decentralized and sensitive data sources, it is increasingly used in many application domains including remote healthcare, smart buildings, and mobile applications. However, FL raises several ethical concerns as it may introduce bias with regard to sensitive attributes (e.g., race, gender, etc.), it is not robust against malicious participants that attempt to poison the data and model, and it remains vulnerable to privacy attacks (e.g., membership inference attacks, etc.). In this talk, we will first discuss the open scientific issues in FL bias, robustness and privacy, before presenting novel FL protocols for handling them.

Biography

Sara Bouchenak is a distinguished Professor of Computer Science at INSA Lyon engineering school. She was awarded the title of Knight in the Order of Academic Palms in 2023. She was the Director of the Federation of Computer Science Research Laboratories in Lyon - Saint Étienne in 2021-2025, encompassing around 800 people. She is the Chair of the Women in Computing Science Committee, at the Department of Computer Science – INSA Lyon. She conducts research on dependable, privacy-preserving and robust distributed systems and distributed/federated learning. She is involved in the organization of many conferences and editorial activities, and serves as general chair (SRDS 2019), program committee chair (e.g., ACM Middleware 2026, etc.), program committee member (e.g., IEEE DSN 2026, IEEE ICDCS 2026, etc.), associate editor (e.g., IEEE Trans. On Dependable and Secure Compuring 2020-2023). Prior to that, Sara was Associate Professor at the University of Grenoble until 2014, a visiting researcher at Universidad Politécnica de Madrid, Spain, in 2009/2010 in the group of Marta Patiño and Ricardo Jimenez, a post-doctoral researcher at EPFL, Switzerland, in 2003 with Willy Zwaenepoel. She received her HDR (Habilitation à Diriger les Recherches) from the University of Grenoble in 2010, and her PhD in computer science from Grenoble Institute of Technology in 2001.

Abstract

PlumTree is an event dissemination algorithm designed in 2007 that aims at balancing the efficiency of structured approaches (such as the use of spanning tree) and the robustness of unstructured approaches (namely, gossip-based broadcast). In this talk we present the background that led to the PlumTree design, how PlumTree meets its goals, and describe the path that led PlumTree to be adopted in the wild, including in application areas that did not exist when PlumTree was designed, from the dissemination of management information of highly distributed key-value stores (RIAK-DB), Blockchain solutions (through the libp2p that uses a publish-subscribe solution based on PlumTree), and more recently Web 3.0 (with its inclusion on the Iroh networking stack for decentralized systems).

Biography

Luís Rodrigues graduated (1986), has a Master (1991) and a PhD (1996) in Electrical and Computer Engineering, by the Instituto Superior Técnico (IST) da Universidade de Lisboa. He obtained the "Agregação" in Informatics (2003) by the Universidade de Lisboa He is a Professor (Professor Catedrático) at Departamento de Engenharia Informática, Instituto Superior Técnico, Universidade de Lisboa. From 1996 to July 2007 he served at the Departmento de Informática, Faculdade de Ciências (Faculty of Sciences), Universidade de Lisboa. He initiated his academic career at the Electrotechnic and Computers Engineering Department of Instituto Superior Técnico de Lisboa (IST) in 1989. From 1986 to 1996 he was a member of the Distributed Systems and Industrial Automation Group at INESC. From 1997-2007, he was a (founding) member of the LASIGE laboratory at University of Lisbon, first as a member of the Navigators group and later as the leader of the Distributed Algorithms and Network Protocols group. He served as Director of the LASIGE in 2004-2005 and he served in the board of directors of INESC-ID Lisboa from 2010-2017. From July 2007 he is a member of the Distributed Systems Group at INESC-ID Lisboa.

João Leitão is an Associate Professor in the Department of Computer Science at Faculdade de Ciências e Tecnologia of the Universidade Nova de Lisboa, and an Integrated Member of the NOVA Laboratory for Computer Science and Informatics (NOVA LINCS). Previous to that, in 2013 João became a Postdoc researcher (and later an integrated member of the NOVA Laboratory for Computer Science and Informatics) of the Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa. In 2015 he became an Invited assistant professor, and in 2016 an Assistant Professor at the Department of Computer Science at FCT - Universidade Nova de Lisboa.

Abstract

Serverless computing is defined by three tenets: autoscaling with scale-to-zero, pay‑as‑you‑go economics, and minimal infrastructure management for users. A decade after the first serverless platforms were deployed, these systems have matured beyond short‑lived, stateless functions to support stateful operation and long‑lived execution. In the era of generative AI, we ask whether the core serverless tenets remain relevant for platforms running AI workloads, and what advantages and pitfalls arise from using AI for platform research. We present our team’s research‑ and engineering‑focused work addressing three key concerns: fast model actuation to reduce vLLM cold‑start latency, improved KV cache performance through user‑level hints called span queries, and a technique termed forward compilation to optimize kernel performance. We also examine how AI‑assisted development has informed our research and present early findings on characterizing evolutionary optimization approaches to better understand the limits of using AI as a research partner.

Biography

Paul Castro, Ph.D. is a Senior Research Manager at the IBM Watson Research Center, where he works on hybrid cloud platforms for generative AI and co-leads the Platform Foundations for AI research theme. His work focuses on applying serverless computing to AI model serving, agentic programming models and runtimes, and cloud‑native developer tooling. A veteran of cloud and mobile systems research, he was an early contributor to Apache OpenWhisk and has led and contributed to widely deployed cloud platforms, earning multiple IBM Research Accomplishment Awards and holds 44 patents. Paul received his Ph.D. in Computer Science from UCLA and brings deep experience at the intersection of systems research, platform engineering, and applied AI.

Abstract

Dune Analytics operates DuneSQL, a data platform for blockchain analytics built on Trino, Delta Lake, and Parquet. Beyond query execution, DuneSQL manages a heterogeneous data lake of public and private tables across dozens of blockchains, with catalog search, a table explorer, team-based access control, billing instrumentation, and REST APIs for both interactive users and programmatic consumers. Platforms of this scope inevitably accumulate operational tasks: metadata indexing, storage cleanup, vacuuming and compaction, search reindexing, billing metering, and data export delivery. In our experience, these tasks grow organically and sprawl across services, repositories, runtimes, and inconsistent approaches to alerting and observability. At Dune, that sprawl grew into more than 50 jobs and services spread across Kubernetes CronJobs and Deployments, Prefect flows, Databricks jobs, and AWS Lambdas, each with its own retry logic and failure modes. This is not unique to Dune; it is a common outcome in growing data platforms when operational concerns are handled locally by the teams closest to them. Left unchecked, it becomes unmanageable. DuneCP is the event-driven control plane we built to address this. In this talk, we describe how DuneCP evolved from a consolidation effort into a system that receives and acts on events signaling maintenance or other internal actions: tracking storage consumption, reindexing table metadata, checking and optimizing table health, purging tombstoned data, and orchestrating incremental table export. It models operations as composable events with lease-based worker coordination, heartbeat liveness, and fault-tolerant recovery. It is used in production for proactive table maintenance, table storage metering, and data delivery to customer Snowflake and BigQuery accounts. We close with lessons from production and ongoing work to expand its scope and flexibility.

Biography

Miguel Mascarenhas Filipe is a Principal Engineer at Dune Analytics, where he leads the Data Warehouse team. He designed DuneCP, an event-driven control plane for DuneSQL - Dune's Trino-based distributed query engine for blockchain data at scale - covering table health management, workload-aware query routing, and cross-platform data delivery orchestration. He previously led the design and development of DuneSQL itself, built on Parquet/Delta Lake over S3 across millions of tables. Before Dune, Miguel was a member of the launch team for Amazon DynamoDB, where he worked on its control plane (partition management, replica balancing, hotspot mitigation, and auto-healing of partitions and servers) and was the main author of its provisioned throughput admission control system. He holds 13 US patents on workload metering and throughput maximization with QoS guarantees. He also held senior engineering roles at Microsoft (Skype/ARIA data pipeline), Citymapper (distributed systems lead), and Unbabel (core platform re-architecture). Miguel holds an MSc in Computer and Software Engineering from Instituto Superior Técnico, Lisbon.