NorCal DB Day 2026

NorCal DB, in a day

NorCalDB Day is a single-day, workshop-style event where participants from academia and industry in Northern California meet to present ideas and discuss their research and experiences.

Fifteen years of NorCal DB

Hosted across the region’s universities and companies since 2011.

Schedule

Accepted Posters

• Interactive Steering of LLM-Powered Scoring Judges with Attune Bhavya Chopra, Meng Chen, Rebecca Dang, Chanbin Park, Shreya Shankar, Sepanta Zeighami, Bjoern Hartmann, Aditya Parameswaran (UC Berkeley)

  Large language models (LLMs) are increasingly used to score text records at scale (e.g., rating candidate resumes on a 1-5 scale). However, existing LLM-powered approaches do not account for the fact that effective scoring requires both holistic understanding of records and locally consistent judgments across similar ones. We present Attune, a mixed-initiative system for steerable LLM-powered scoring. Given a task description and scoring range, Attune performs pairwise comparisons across records to develop a global understanding first, and then resolves these comparisons into consistent score assignments—deriving scoring criteria and rules bottom-up in the process. These serve as shared representations of scoring logic that users can inspect and edit. Based on insights from a formative study (n = 12), Attune’s interface introduces novel steering interactions that allow users to deterministically refine scoring logic. Users can provide examples, directly edit criteria, rules, or target distributions, and give natural language feedback—with all refinements compiling into constraints that guide re-scoring. We validate our approach through a technical evaluation across three workloads and a user study with domain experts (n = 8) in healthcare, law, education, and AI evaluation.

• Meaningful Data Deletion in Data Processing Pipelines Arnav Dhariya, Vishal Chakraborty, Sharad Mehrotra (UC Irvine)

  Modern systems fail to provide privacy guarantees for deleted data. Meaningful deletion mandates a guarantee that limits what can be inferred about a deleted value from a post-deletion database state. Prior work introduced the P2E2 equivalence, a deterministic framework that achieves meaningful erasure by deleting only enough to reduce inferential power to what was possible before insertion of the target cell. However, eliminating every inference path results in excessive deletions of dependent cells, effectively leaving the data unusable.

  This motivates a probabilistic relaxation of meaningful deletion: rather than eliminating all inference paths, bounding re-inference probability to a leakage budget permits far fewer auxiliary deletions. We capture data dependencies as weighted relational dependency rules (RDRs) and realize deletion through two mechanisms: Exp, which enumerates all possible masks, and Gum, which uses the Gumbel-max trick. Both methods reduce deletions while controlling leakage through deletion pattern and surviving data. We evaluate them across five datasets and achieve substantial deletion reduction compared to P2E2 while respecting the leakage budget. We extend this framework beyond static instances to data processing pipelines where node-specific corrective actions are required to ensure end-to-end meaningful deletion.

• Prela: A composable, controllable query language Remy Yisu Wang, Paul Talma (UCLA)

  We revisit Tarski’s Algebra of Relations (TAR), an old formalism of relations predating Codd’s relational algebra by over 100 years, as a new foundation for relational databases. We argue TAR provides a better abstraction at both the semantic level and the physical level, in the context of modern application code and system architecture. To demonstrate the strengths of TAR, we design and implement Prela, a compositional and controllable query language, and show that queries written in Prela are concise, clear, and efficient.

• Optimal Optimization of Optimal Joins Zheng Luo (UCLA), Wim Van den Broeck (University of Bergen), Sean Hoffmeister (UCLA), Remy Wang (UCLA)

  There has been renewed interest in instance-optimal join algorithms for acyclic queries, as well as efforts to make them more practical. Execution of these algorithms is typically guided by a join tree. A query may have many different join trees, and any of them guarantees the optimal asymptotic complexity, yet choosing a good tree remains crucial for practical performance. In this paper, we present O3, a linear-time algorithm that constructs an optimal join tree for any alpha-acyclic query. We implement our algorithm in DuckDB, and compare it with prior algorithms on standard benchmarks containing small queries, as well as synthetic ones spanning millions of tables. The experiments show our algorithm matches the end-to-end performance of the baselines on small queries, while gaining orders-of-magnitude speedup on the large ones.

• Database Views as Explanations for Relational Deep Learning Agapi Rissaki (Northeastern University), Ilias Fountalis (RelationalAI), Wolfgang Gatterbauer (Northeastern University), Benny Kimelfeld (Technion & RelationalAI)

  In recent years, there has been significant progress in the development of deep learning models over relational databases, including architectures based on heterogeneous graph neural networks (hetero-GNNs) and heterogeneous graph transformers. In effect, such architectures state how the database records and links (e.g., foreign-key references) translate into a large, complex numerical expression, involving numerous learnable parameters. This complexity makes it hard to explain, in human-understandable terms, how a model uses the available data to arrive at a given prediction. We present a novel framework for explaining machine-learning models over relational databases, where explanations are view definitions that highlight focused parts of the database that mostly contribute to the model’s prediction. We establish such global abductive explanations by adapting the classic notion of determinacy by Nash, Segoufin, and Vianu (2010). In addition to tuning the tradeoff between determinacy and conciseness, the framework allows controlling the level of granularity by adopting different fragments of view definitions, such as ones highlighting whole columns, foreign keys between tables, relevant groups of tuples, and so on. We investigate the realization of the framework in the case of hetero-GNNs, and develop a model-specific approach via the notion of learnable masks. For comparison, we propose model-agnostic heuristic baselines and show that our approach is both more efficient and achieves better explanation quality in most cases. Our extensive empirical evaluation on the RelBench collection across diverse domains and record-level tasks demonstrates both the usefulness of our explanations and the efficiency of their generation.

• Manticore: Accelerated Semantic Concurrency Control for Snapshot Transactions in a Modern Document Database William Schultz (MongoDB), Saikrishna Achalla (UCLA), Remy Wang (UCLA), Murat Demirbas (MongoDB)

  Most modern transactional concurrency control algorithms target their focus at the key-value abstraction layer, both defining isolation correctness and validating transactions at this low level of detail. Particularly for optimistic concurrency control schemes, this leads to poor performance under high contention due to write conflicts between concurrent transactions. Reasoning at the key-value storage layer also destroys much of the rich, expressive semantic information that exists at the application layer once it is translated to storage operations. In this work, we propose a optimistic semantic-oriented concurrency control engine that operates at the query processing layer and takes advantage of the concrete semantics of application level update operations in order to achieve zero-percent abort rates. Under certain workloads, we observe throughput improvement under increasing contention levels, with up to 20x throughput gains under the highest contention levels compared to traditional concurrency control mechanisms.

• HashBound: Fast and Accurate Cardinality Upper Bounds for Acyclic Queries Chun Wang, Nikolaos Tziavelis (UC Santa Cruz)

  We present HashBound, a guaranteed upper-bound cardinality estimator specialized for acyclic join queries. HashBound partitions the join-key domain into k buckets, storing simple sum and max degree statistics per bucket to compress each attribute into k entries. At query time, the upper bound is computed via a single traversal of the join tree where each edge requires O(k) computation—yielding a process reminiscent of a compressed Yannakakis counting algorithm. Rather than relying on a random hash function, we employ lightweight learning via k-means clustering during preprocessing. By representing each join-key by its degree profile and grouping structurally similar keys together, HashBound improves the error introduced by per-bucket approximations. In contrast to state-of-the-art bounding methods that solve a global linear program per query, HashBound’s estimation is solver-free and proceeds from smaller subquery estimates to larger ones.

• A Learned Query Optimizer for Filtered ANN Search Abylay Amanbayev, Tri Dang, John Cedric Alora, Florin Rusu (UC Merced)

  Filtered approximate nearest-neighbor (ANN) search is a per-query optimization problem: the lowest-latency execution plan satisfying a recall requirement depends on the query, its filter, and the available indexes. ANN indexes also expose query-time hyperparameters that control the search budget and are typically tuned manually and fixed across a workload. We present a compact learned optimizer that jointly selects an execution plan and a query-specific hyperparameter value using features available before vector search: filter selectivity and a cheap estimate of Global–Local Selectivity (GLS) correlation. Varying an internal safety margin traces a throughput–recall-safety frontier, reserving expensive search for difficult queries. Across four hard and superhard MoReVec workloads, when 95% of queries meet a Recall@10 target of 0.95, the optimizer achieves 1.7–2.6× the search throughput available on the workload-wide static index–parameter frontier at the same success rate.

• Pruning as You Go: Metadata-Only Join Graph Inference Tri Dang, Abylay Amanbayev, Florin Rusu (UC Merced)

  AI-driven workflows over relational data, such as Text-to-SQL, depend on join relationships that are often missing, incomplete, or unreliable in real-world schemas. We study metadata-only join graph inference, where attribute-level join relationships must be recovered from table names, column names, data types, and limited key information without access to data values. Existing approaches verify candidate pairs independently over a fixed candidate set, overlooking the fact that accepted relationships introduce structural constraints that can eliminate competing candidates and reduce the remaining search space. Verification order therefore determines not only the cost of each decision but also which decisions remain. We propose Adaptive Cluster-Guided Join Discovery (ACJD), a framework that treats join inference as a stateful search process. ACJD organizes unresolved candidates into semantic cluster-pair regions, prioritizes candidates using semantic-key similarity, and applies relationship-aware pruning immediately after relationship acceptance. Regions that yield no positive evidence terminate early, avoiding unnecessary verification. Experiments on TPC-H, TPC-DS, BIRD-SQL, and JOB with a GPT-5.5 verifier show that adaptive exploration and the acceptance-time pruning it drives reduce LLM verification workload by 58–78% compared with exhaustive verification of the same candidate space while maintaining comparable or better inference quality. Against direct pairwise verification of all metadata-pruned candidates, the full pipeline matches F1 on three benchmarks and improves F1 on BIRD-SQL while requiring only 17–31% as many candidate evaluations and achieving 4–11× lower API costs. Compared with Nexus, the closest metadata-only baseline, ACJD improves F1 on both BIRD-SQL and TPC-DS under a shared GPT-4o backend.

• Direct Access to Query Answers Jiayin Hu, Nikolaos Tziavelis (UC Santa Cruz)

  Direct access asks for the retrieval of query answers by their ranked position, given a query and a desired order. While the time complexity of data structures supporting such accesses has been studied in depth, and efficient algorithms for many queries and common orders are known, their practical performance has received little attention. We provide an implementation covering a wide range of queries and orders; it allows us to investigate intriguing practical aspects, including the comparative performance of database systems and the relationship between direct access and its single-access counterpart.

• MemOS: Wide Sharing for Data Intensive Applications Daniel Bittman (Elephance Memory), Pankaj Mehra (Elephance Memory and OSU)

  Data intensive applications such as AI demand higher performance, parallelism, and data capacity from hardware, programming models, and infrastructure-level software. Memory-semantic fabrics (CXL, UALink, etc.) enable access to fabric-attached memory, allowing applications to scale their data while exposing data parallelism. Meanwhile, programmers are stuck with host-centric virtual memory and compute-centric programming models which limit the ease and efficiency with which applications can move and manipulate data in dynamic, multi-tiered memory hierarchies, or share memory on fabric-attached devices among cooperating nodes and applications. Further, applications often must construct and destruct many runtime data structures, particularly when switching models or source data, that are not directly sharable across address spaces. Elephance introduces the Memory Objects abstraction for zero-overhead fluid movement and sharing of even complex data structures across heterogeneous systems and memory tiers. Memory objects provide movable and sharable in-memory data structures with memory access semantics designed and optimized for data-centric, in-memory processing, including expressing logical data locality, and reducing page-faults and TLB operations. We show that memory objects enable dramatic reductions in model load time (>100x), switch time (4x), and peak memory usage (9x) for several models on a modified llama.cpp, while demonstrating wide-sharing for key model pipeline data structures. These results extend to other AI frameworks and generalize to other data intensive applications, enabling wide sharing, reduced memory usage, and improved parallelism for complex data structures.

• Temporal Motif Estimation using Spanning Tree Sampling Yunjie Pan (Meta), Omkar Bhalerao (UC Santa Cruz), C. Seshadhri (UC Santa Cruz), Nishil Talati (University of Illinois Urbana-Champaign)

  The mining of pattern subgraphs, known as motifs, is a core task in the field of graph mining. Edges in real-world networks often have timestamps, so there is a need for temporal motif mining. A temporal motif is a richer structure that imposes timing constraints on the edges of the motif. Temporal motifs have been used to analyze social networks, financial transactions, and biological networks.

  Motif counting in temporal graphs is particularly challenging. A graph with millions of edges can have trillions of temporal motifs, since the same edge can occur with multiple timestamps. There is a combinatorial explosion of possibilities, and state-of-the-art algorithms cannot manage motifs with more than four vertices.

  In this work, we present TIMEST: a general, fast, and accurate estimation algorithm to count temporal motifs of arbitrary sizes in temporal networks. Our approach introduces a temporal spanning tree sampler that leverages weighted sampling to generate substructures of target temporal motifs. This method carefully takes a subset of temporal constraints of the motif that can be jointly and efficiently sampled. TIMEST uses randomized estimation techniques to obtain accurate estimates of motif counts.

  We give theoretical guarantees on the running time and approximation guarantees of TIMEST. We perform an extensive experimental evaluation and show that TIMEST is both faster and more accurate than previous algorithms. Our CPU implementation exhibits an average speedup of 28x over state-of-the-art GPU implementation of the exact algorithm, and 6x speedup over SOTA approximate algorithms while consistently showcasing less than 5% error in most cases. For example, TIMEST can count the number of instances of a financial fraud temporal motif in four minutes with 0.6% error, while exact methods take more than two days.

• Oblivious Computation at Terabyte Scale: Relational Databases Apostolos Mavrogiannakis (UC Santa Cruz)

  Trusted execution environments let clients outsource computation to untrusted clouds, but they do not hide memory access patterns, which leakage-abuse and side-channel attacks exploit. Oblivious algorithms close this gap, yet they have struggled to meet the performance and scale of real workloads: until recently, the best oblivious joins were evaluated on megabytes of data. This talk follows two systems that push that limit. Obliviator targets a single shared-memory enclave. It introduces an oblivious aggregation tree and a sort-free parallel expansion algorithm, cuts the oblivious sorts in a non-foreign-key join from five to two, and parallelizes every step, running up to 80× faster than the prior state of the art and scaling to hundreds of gigabytes. DISCO then moves beyond one machine, where communication patterns must be hidden too. By partitioning data vertically instead of horizontally, it lets servers compute in parallel with minimal coordination, outperforms prior distributed oblivious joins by up to 180×, and reports the first results at terabyte scale, up to 16 TB. I close with open problems in oblivious query processing, including complex multi-join queries.

• Blockchain RPC Providers See Every Query: A Leakage Analysis Sandhya Saravanan (UC Santa Cruz)

Three ways in

Key deadlines

• September 15Poster submission closes (title & short abstract)
• September 20Poster notifications
• October 2Registration closes
• NorCal DB Day 2026

Organizing committee

Questions? Write to .