Poster 1: ANN-Benchmarks Extension for Filtered Search
Authors: Abylay Amanbayev, Brian Tsan, Florin Rusu, UC Merced
Abstract: Retrieval-Augmented Generation (RAG) applications increasingly rely on Approximate Nearest Neighbor Search (ANNS) over text embeddings, often combined with metadata filtering to exclude irrelevant results. Although ANNS has been extensively studied, there is little work analyzing workloads with filters, and limited datasets with metadata attributes and vectors based on text embeddings. To this end, we extend ANN-Benchmarks to support queries with metadata filters, analyzing effects of filter selectivity and search parameters on query performance. We also introduce datasets derived from movie synopses (over 500,000 vectors) and user reviews (over 2 million vectors) on IMDb, enriched with over 5 metadata attributes per vector, and corresponding query workloads. Extensive experiments across three open-source vector search systems – Milvus, pgvector, and FAISS – reveal some intriguing initial results. Additionally, we implement support for predicate evaluation in the HNSW index in FAISS, ensuring comparable performance evaluation with other systems.
Poster 2: HotStuff-1: Linear Consensus with One-Phase Speculation
Presenter: Dakai Kang, UC Davis
Paper Link: https://dl.acm.org/doi/10.1145/3725308
Abstract: This paper introduces HotStuff-1, a BFT consensus protocol that improves the latency of HotStuff-1 by two network hops while maintaining linear communication complexity against faults. Furthermore, HotStuff-1 incorporates an incentive-compatible leader rotation design that motivates leaders to propose transactions promptly. HotStuff-1 achieves a reduction of two network hops by speculatively sending clients early finality confirmations, after one phase of the protocol. Introducing speculation into streamlined protocols is challenging because, unlike stable-leader protocols, these protocols cannot stop the consensus and recover from failures. Thus, we identify prefix speculation dilemma in the context of streamlined protocols; HotStuff-1 is the first streamlined protocol to resolve it. HotStuff-1 embodies an additional mechanism, slotting, that thwarts delays caused by (1) rationally-incentivized leaders and (2) malicious leaders inclined to sabotage others' progress. The slotting mechanism allows leaders to dynamically drive as many decisions as allowed by network transmission delays before view timers expire, thus mitigating both threats.
Poster 3: FairDAG: Consensus Fairness over Multi-Proposer Causal Design
Presenter: Dakai Kang, UC Davis
Paper Link: https://arxiv.org/abs/2504.02194
Abstract: The rise of cryptocurrencies like Bitcoin and Ethereum has driven interest in blockchain database technology, with smart contracts enabling the growth of decentralized finance (DeFi). However, research has shown that adversaries exploit transaction ordering to extract profits through attacks like front-running, sandwich attacks, and liquidation manipulation. This issue affects blockchain databases in which block proposers have full control over transaction ordering. To address this, a more fair approach to transaction ordering is essential.
Existing fairness protocols, such as Pompe and Themis, operate on leader-based consensus protocols, which not only suffer from low throughput, but also allow adversaries to manipulate transaction ordering. To address these limitations, we propose FairDAG-AB and FairDAG-RL that run fairness protocols on top of DAG-based consensus protocols, which improve protocol performance in both throughput and fairness quality, leveraging the multi-proposer design and validity of DAG-based consensus protocols.
We conducted a comprehensive analytical and experimental evaluation of our protocols. The results show that FairDAG-AB and FairDAG-RL outperform the prior fairness protocols in both throughput and fairness quality.
Poster 4: Carry the Tail in Consensus Protocols
Presenter: Dakai Kang, UC Davis
Paper Link: https://arxiv.org/pdf/2508.12173
Abstract: We present Carry-the-Tail, the first deterministic atomic broadcast protocol in partial synchrony that, after GST, simultaneously guarantees two desirable properties: (i) a constant fraction of commits are proposed by non-faulty leaders against tail-forking attacks, and (ii) optimal, worst-case quadratic communication under a cascade of faulty leaders. The solution also guarantees linear amortized communication, i.e., the steady-state is linear. Combining these two desirable properties was not simultaneously achieved previously: on one hand, prior atomic broadcast solutions achieve per-view linear word communication complexity. However, they face a significant degradation in throughput
under tail-forking attack. On the other hand, existing solutions to tail-forking attacks require either quadratic communication steps or computationally-prohibitive SNARK generation.
The key technical contribution is Carry, a practical drop-in mechanism for streamlined protocols in the HotStuff family. Carry guarantees good performance against tail-forking and removes most leader-induced stalls, while retaining linear traffic and protocol simplicity. Carry-the-Tail implements the Carry mechanism on HotStuff-2.
Poster 5: Execution to Explanation: Profiling ResilientDB with AI
Presenter: Bismanpal Anand, UC Davis
Abstract: Flamegraphs can do more than reveal bottlenecks—they can teach us how systems execute. With ResLens, we continuously profile ResilientDB, capturing userspace and system calls and transforming them into interactive flamegraphs that narrate execution paths. These visualizations, paired with disk metrics, blockchain state, and live PBFT views, offer a rich lens into system design and behavior. Building on this foundation, ResAI introduces an AI-assisted exploration layer: answering natural language questions, surfacing performance insights, and even generating and testing smart contracts. By combining continuous profiling with AI-guided interaction, ResLens and ResAI together transform ResilientDB into a transparent, explainable, and approachable platform for learning, debugging, and advancing distributed systems.
Poster 6: Pushing the Frontiers of Rotational BFT Consensus: Speculation, Tail-Forking Resilience, and Rapid View Synchronization
Presenter: Bismanpal Anand, UC Davis
Abstract: This Gong Show session presents three complementary advances in rotational Byzantine Fault Tolerant (BFT) consensus. HotStuff-1 introduces one-phase speculation with a novel slotting mechanism, cutting latency by two network hops while maintaining linear communication and mitigating both rational and malicious leader delays. Carry-the-Tail addresses tail-forking attacks by embedding the Carry mechanism into HotStuff-2, ensuring resilience with linear amortized communication complexity. SpotLess pioneers a concurrent rotational consensus design combined with Rapid View Synchronization, eliminating a separate complex view synchronization mechanism. Together, these works advance rotational consensus protocols toward practical, high-performance, and attack-resilient blockchain systems.
Poster 7: Scalable Censorship-Resistant DAG Consensus via Trusted Components
Authors: Shaokang Xie, Dakai Kang, Hanzheng Lyu, Jianyu Niu, Mohammad Sadoghi, UC Davis
Presenter: Shaokang Xie, University of California, Davis
Abstract: We introduce Fides, an asynchronous DAG-based BFT protocol that uses TEEs to (1) reduce the quorum size from n=3f+1 to n=2f+1, (2) cut communication to linear complexity with light cryptographic reliance (avoiding heavyweight global coins), and (3) provide low-cost, guaranteed censorship resilience. Fides isolates four small trusted components inside TEEs to keep the TCB minimal: Reliable Broadcast, Vertex Validation, Common Coin, and Transaction Disclosure. The extensive experiments on geo-distributed and local clusters show Fides outperforming other state-of-the-art protocols.
Poster 8: Sketched Sum-Product Networks for Joins
Presenter: Brian Tsan, UC Merced
Abstract: Sketches can achieve highly accurate join cardinality estimation, a critical task in query optimization. Accurately estimating cardinality — analogous to computational cost — allows optimizing query execution costs in database systems. However, sketches are typically constructed for predefined query selections known in advance, limiting their applicability to new queries. To address this, we propose using Sum-Product Networks (SPNs) to approximate sketches dynamically. SPNs model multivariate distributions, such as relations, as linear combinations of univariate distributions. By representing univariate distributions as sketches, SPNs can efficiently combine them element-wise to approximate the sketch of any selection on-the-fly. We implement the Fast-AGMS and Bound Sketch methods, which have demonstrated success in prior work, despite high construction costs. By approximating them with SPNs, we offer a more scalable solution for improved query optimization.
Poster 9: Toward Building Efficient Document Analytics Systems from the Lens of Document Structure
Presenter: Yiming Lin, UC Berkeley
Paper links: https://yiminglin18.com/publication/zendb/, https://yiminglin18.com/publication/twix/
Abstract: The vast majority—over 80%—of data today exists in unstructured formats, and querying and extracting value from unstructured document collections remains a considerable challenge. Although current LLM-powered systems support analytical queries over documents, they often reduce documents to bags of words, ignoring their semantics and structure. This poster presents a series of works that explore diverse structures hidden within document collections and demonstrates that uncovering these structures can effectively facilitate multiple document analysis tasks. At one extreme, we explore documents sharing a similar high-level template that impart a common semantic structure, such as scientific papers from the same venue. We introduce ZenDB, a document analytics system that leverages this semantic structure, coupled with LLMs, to answer ad-hoc SQL queries on document collections. At another extreme, we explore documents that are programmatically generated by populating fields in a visual blueprint, such as invoices, order bills, containing structured data like tables or key-value pairs. We present TWIX, a document analytics tool that first infers the common blueprint and then extracts structured data from documents efficiently. Finally, we explore another type of document with loose structure, where only metadata (e.g., headers) are available, and introduce LSF, a technique that scales document analysis by modeling common loose structures across documents.
Poster 10: Steering Data Processing with DocWrangler
Presenter: Shreya Shankar, UC Berkeley
Paper links: https://dl.acm.org/doi/abs/10.14778/3746405.3746426, https://dl.acm.org/doi/10.1145/3746059.3747625
Abstract: Unstructured text has long been difficult to automatically analyze at scale. Large language models (LLMs) now offer a way forward by enabling semantic data processing, where familiar data processing operators (e.g., map, reduce, filter) are powered by LLMs instead of code. However, building effective semantic data processing pipelines presents a departure from traditional data pipelines: users need to understand their data to write effective pipelines, yet they need to construct pipelines to extract the data necessary for that understanding -- all while navigating LLM idiosyncrasies and inconsistencies. We present DocWrangler, a mixed-initiative integrated development environment (IDE) for semantic data processing with three novel features to address the gaps between the user, their data, and their pipeline: (i) In-Situ User Notes that allows users to inspect, annotate, and track observations across documents and LLM outputs, (ii) LLM-Assisted Prompt Refinement that transforms user notes into improved operations, and (iii) LLM-Assisted Operation Decomposition that identifies when operations or documents are too complex for the LLM to correctly process and suggests decompositions. Our evaluation combines a think-aloud study with 10 participants and a public-facing deployment with 1,500+ recorded sessions, revealing how users develop systematic strategies for their semantic data processing tasks; e.g., transforming open-ended operations into classifiers for easier validation and intentionally using vague prompts to learn more about their data or LLM capabilities.
Poster 11: Cut Costs, Not Accuracy: LLM-Powered Data Processing with Guarantees
Presenter: Sepanta Zeighami
Paper links: https://arxiv.org/pdf/2509.02896
Abstract: Large Language Models (LLMs) are being increasingly used as a building block in data systems to process large text datasets. To do so, LLM model providers offer multiple LLMs with different sizes, spanning various cost-quality trade-offs when processing text at scale. Top-of-the-line LLMs (e.g., GPT-4o, Claude Sonnet) operate with high accuracy but are prohibitively expensive when processing many records. To avoid high costs, more affordable but lower quality LLMs (e.g., GPT-4o-mini, Claude Haiku) can be used to process records, but we need to ensure that the overall accuracy does not deviate substantially from that of the top-of-the-line LLMs. The model cascade framework provides a blueprint to manage this trade-off, by using the confidence of LLMs in their output (e.g., log probabilities) to decide on which records to use the affordable LLM. However, existing solutions following this framework provide only marginal cost savings and weak theoretical guarantees because of poor estimation of the quality of the affordable LLM’s outputs. We present BARGAIN, a method that judiciously uses affordable LLMs in data processing to significantly reduce cost while providing strong theoretical guarantees on the solution quality. BARGAIN employs a novel adaptive sampling strategy and statistical estimation procedure that uses data and task characteristics and builds on recent statistical tools to make accurate estimations with tight theoretical guarantees. Variants of BARGAIN can support guarantees on accuracy, precision, or recall of the output. Experimental results across 8 real-world datasets show that BARGAIN reduces cost, on average, by up to 86% more than state-of-the-art, while providing stronger theoretical guarantees on accuracy of output, with similar gains when guaranteeing a desired level of precision or recall.
Poster 12: Rethinking LLM-Powered Dataset Discovery with DataScout
Presenter: Bhavya Chopra
Abstract: Dataset Search -- the process of finding appropriate datasets for a given task -- remains a critical yet under-explored challenge in data science workflows. Assessing dataset suitability for a task (e.g., training a classification model) is a multi-pronged affair that involves understanding: data characteristics (e.g. granularity, attributes, size), semantics (e.g., data semantics, creation goals), and relevance to the task at hand. Present-day dataset search interfaces are restrictive -- users struggle to convey implicit preferences and lack visibility into the search space and result inclusion criteria -- making query iteration challenging. To bridge these gaps, we introduce DataScout to proactively steer users through the process of dataset discovery via -- (i) AI-assisted query reformulations informed by the underlying search space, (ii) semantic search and filtering based on dataset content, including attributes (columns) and granularity (rows), and (iii) dataset relevance indicators, generated dynamically based on the user-specified task. A within-subjects study with 12 participants comparing DataScout to keyword and semantic dataset search reveals that users uniquely employ DataScout's features not only for structured explorations, but also to glean feedback on their search queries and build conceptual models of the search space.
Poster 13: Visual Relevancy Index System for Object Localization in Video Query Processing
Presenter: Chanwut Kittivorawong
Poster: https://docs.google.com/presentation/d/1feZrDl45pZvS_D4kWwi-g93rXlFaVb76_B9rhs3ocBE
Abstract: Video query processing is essential for a wide range of analytics tasks, but the computational cost of running machine learning (ML) models on large video datasets is a major bottleneck. Video data exhibits high degrees of visual and temporal redundancy, yet existing video processing systems often fail to exploit these characteristics fully. Current approaches often treat ML models as black boxes, limiting optimizations to the frame level by, for example, reducing the frame rate. This creates an undesirable trade-off between processing speed and the fineness of object tracking. Our work addresses this limitation by leveraging the visual redundancy within each frame. We propose a visual redundancy index for object localization and tracking that identifies and removes irrelevant pixel content from video frames before they are input into ML models. This reduces the ML model workload and decouples processing speed from tracking fineness, enabling more efficient video analytics with an alternative trade-off in accuracy reduction.