Demos accepted to WoAIS1 (First International Workshop on AI and Serverless, 2026). See the program for the full schedule.
In this demo, we present Blocks-DB, a lightweight stateless serverless vector database designed entirely on top of cloud object storage and Function-as-a-Service (FaaS) primitives. Blocks-DB introduces a block-based partitioning strategy that continuously aggregates incoming vectors into balanced immutable blocks, enabling efficient indexing and scalable parallel query execution without relying on persistent servers or complex clustering mechanisms.
Blocks-DB uses a fully event-driven architecture where ingestion, partitioning, indexing, and querying are executed through stateless cloud functions. By decoupling ingestion from indexing and leveraging asynchronous block creation, the system supports continuous ingestion while maintaining stable query latency and recall with low infrastructure overhead.
We have evaluated Blocks-DB against Amazon S3 Vectors, the current state-of-the-art serverless vector storage solution, showing that block-based partitioning improves elasticity and cost efficiency while achieving competitive query latency and recall.
During the demo, we will showcase the Blocks-DB workflow, from system deployment and database creation to vector ingestion, indexing, querying, and GET and PUT operations, demonstrating how lightweight stateless serverless architectures can efficiently support modern vector search workloads.
Octavio Iacoponelli (Universitat Rovira i Virgili)
The rapid advancement of Artificial Intelligence (AI) and data-intensive research has created a profound infrastructure bottleneck. While Python has become the lingua franca of data science, it was not natively designed for cloud-scale, distributed execution. Consequently, researchers, data scientists, and students pay a steep “Cloud Complexity Tax,” spending up to 60% of their time navigating infrastructure provisioning, environment dependencies, and distributed frameworks. To address this educational and productivity gap, we introduce PyRun Cloud: an AI-driven, multi-cloud platform designed to democratize access to distributed supercomputing. Emerging from EU-funded research, PyRun abstracts cloud complexities, allowing users to seamlessly execute high-performance workloads via frameworks like Lithops (FaaS) and Dask (Clusters) with zero manual infrastructure configuration. Recently pivoting toward an academic paradigm, PyRun integrates seamlessly with educational programs like AWS Academy to provide zero-cost infrastructure for universities. Through its novel Learn Lab framework, PyRun offers AI-guided, hands-on experiments—from processing massive genomic datasets via zero-data-movement strategies to distributed geospatial climate analysis. Verifiable benchmarks demonstrate that PyRun's deep-tech core outperforms commercial alternatives, executing Serverless/FaaS workloads up to 5.4x faster and 14.9x cheaper. By unifying AI Copilots, Model Context Protocol (MCP) Agents, and automated runtime management, PyRun enables researchers and students to focus entirely on algorithmic discovery, effectively eliminating the barriers to cloud-scale innovation.
Video [link]
Daniel Alejandro Coll Tejeda (Universitat Rovira i Virgili)
The Function-as-a-Service (FaaS) paradigm is increasingly adopted across the Edge–Cloud Continuum. Modern Edge infrastructures are inherently heterogeneous, often combining ARM and x86 processors with different performance, energy, and cost characteristics. While prior work has shown that architecture-aware scheduling can improve efficiency, most FaaS platforms still lack adaptive mechanisms to exploit heterogeneous resources, particularly when handling previously unseen functions whose architecture affinity is unknown.
This work extends Serverledge, an open-source FaaS framework for the Edge–Cloud Continuum, to support heterogeneous execution environments and architecture-aware scheduling. The framework is enhanced to manage architecture-specific node information and execute functions transparently across ARM and x86 nodes.
To exploit architecture affinity, we formulate architecture selection as a Multi-Armed Bandit (MAB) problem, where the scheduler learns online which architecture best suits each function. We investigate both context-free and contextual MAB algorithms, leveraging runtime information such as memory utilization to guide decisions. The learning component is integrated with a consistent hashing mechanism that distributes requests across nodes while preserving execution locality.
Experimental results on hybrid ARM/x86 clusters show that the proposed approach effectively learns architecture affinity and outperforms static scheduling policies with negligible overhead.
Gabriele Russo Russo (University of Rome Tor Vergata)