A meta-subgraph ( M ), defined as an edge-unfolded subgraph consisting of a set of connected edge types in the schema of a heterogeneous information network (HIN), generalises the classical meta-path. Although prior studies have leveraged instances of special shapes defined on the schema for tasks such as cohesive subgraph discovery, similarity measurement and recommendation, no existing work has formally defined the M -instance search problem and developed efficient algorithms dedicated to it. In this paper, we first generalise M queries and then systematically explore efficient algorithms dedicated to M instance search. For a given query M , the vertex/edge correspondences are fixed by the schema. Thus, we do not search for embeddings; instead, we directly retrieve the corresponding instance subgraphs. This avoids generating embedding permutations, although the decision version remains NP-complete. We propose two baselines adapted from backtracking and degeneracy-ordering strategies, which scale only to small HINs. We then derive a new upper bound on the number of M instances, | Ans | * edge , a specialisation of the AGM bound, which exploits edge-type adjacency in M and typed edge lists. Guided by the idea of the bound, we enhance backtracking with effective pruning, achieving complexity O * (| Ans | * edge ). Extending and generalising this analysis to subgraphs of M adjacency and instances of subgraphs of M yields | Ans | * bag . By fully considering the trade-off over: 1) the cost of deriving optimum subqueries, 2) the accuracy and cost of estimating the number of instances of subqueries, and 3) the cost of materialising instances of subqueries, we propose to utilise star shape subgraphs of M to search M instances, which runs in O * (| Ans | * star ) and is practically much faster due to cross-star based pruning. Experiments on large real-world HINs demonstrate that our best algorithm runs up to two orders of magnitude faster than the baselines and direct AGM-based approaches.
Dynamic graphs, pivotal in applications ranging from social networks to biological systems, pose significant challenges in storage and update efficiency due to their continuous evolution through insertions and deletions. Traditional graph compression methods, primarily optimized for static graphs, often struggle in dynamic environments, leading to costly decompression-recompression cycles during updates. To address this limitation, we propose a novel theoretical framework that enables efficient direct updates on rule-based compressed graphs. Additionally, we introduce a dynamic graph processing framework that balances space efficiency, update responsiveness, and query performance. Our framework treats updates as lightweight, localized modifications, sustains compression via background cleanup, and preserves query performance by ensuring structural integrity. Our method achieves significant memory savings, reducing usage by an average of 50.1% over state-of-the-art dynamic compressed graph systems, while demonstrating highly competitive update throughput and query performance. These advancements establish a scalable, high-performance solution for dynamic graph management, effectively bridging the gap between space efficiency, updates responsiveness and query performance.
Ever since the Dennard scaling broke down in the early 2000s and the frequency of the CPUs stalled, vendors have started to increase the core count in each CPU chip at the expense of introducing heterogeneity, thus ushering the era of NUMA and Chiplet processors. Since then, the heterogeneity in the design space of hardware has only increased to the point that DBMS performance may vary significantly up to an order of magnitude in modern servers. An important factor that affects performance includes the location of the logical cores where the DBMS queries execute, and the location where the data resides. This paper introduces P-MOSS, a learned spatial scheduling framework that schedules query execution to specific logical cores, and co-locates data on the corresponding NUMA node. For cross-hardware and workload adaptability, P-MOSS leverages core principles from Large Language Models, such as Next Token prediction, Generative Pre-training, and Fine-tuning. In the spirit of hardware-software synergy, P-MOSS guides its scheduling decision solely based on the low-level hardware statistics collected from the hardware Performance Monitoring Unit with the aid of a Decision Transformer. Experimental evaluation is performed in the context of the B+-Tree index. Performance results demonstrate that P-MOSS offers an improvement of up to 6 & times; over traditional schedules in terms of query throughput.
Federated learning (FL) has emerged as a popular paradigm for distributed machine learning over decentralized data. A typical FL training task involves a fleet of client devices with private data and a centralized server for aggregating the global model. Data generated by FL clients, e.g., smart phones, vehicles, and cameras, is prone to noise. While the impact of data noise on centralized learning (CL) is well understood, to our best knowledge there is a lack of a systematic study from this point of view for FL. In this paper, we fill this gap by presenting an empirical investigation to provide a deeper understanding regarding the impact of data noise on FL. Our study is enabled by DataNoiseGenerator, an open-source and extensible toolkit that we developed for the injection of controlled data noise across five diverse data modalities: image, video, audio, text, and tabular data. We then carry out extensive experiments based on the noisy data generated by DataNoiseGenerator, and our experimental evaluation results reveal that FL is significantly more vulnerable to data noise compared to CL, in terms of the quality of the trained ML models. This gap between FL and CL widens as the intensity of data noise and the proportion of noisy FL clients increase. We further present a detailed analysis to diagnose the root cause of this increased sensitivity of FL to data noise. Our analysis finds that the aggregation performed by the FL server can amplify divergent updates from FL clients trained on noisy data, thereby hindering global model convergence. We conclude that data quality issues are a fundamental challenge for deploying robust FL systems and demand novel decentralized data cleaning mechanisms.
We introduce De Finetti Logic, a novel formalization of the data model of Gamma Probabilistic Databases. De Finetti Logic is a simple, database-centric probabilistic programming framework in which models are specified through relational constraints applied to a probabilistic database. This framework is grounded in the concept of Polya urns and offers an intuitive formalization of exchangeability, the fundamental structural assumption underlying Gamma Probabilistic Databases. We leverage this new formalism to develop a novel inference mechanism based on variational methods. Starting from a DFL theory, our inference method automatically identifies an appropriate family of variational distributions to approximate the target posterior, and synthesizes an optimization algorithm to minimize the Kullback-Leibler divergence between the true posterior and its variational surrogate. To maximize performance, we employ knowledge compilation techniques to limit the number of variational parameters, while preserving the expressive power of the surrogate. Our method is implemented as an extension to the StarfishDB system. Experimental results demonstrate that (i) our approach provides a practical, scalable, and fast-converging alternative to general-purpose inference via collapsed Gibbs sampling, and (ii) remains competitive with specialized algorithms that leverage model-specific optimizations.
Subgraph-based graph representation learning (SGRL) is an emerging class of GNN models that achieve much higher accuracy for various tasks than classical GNN models (e.g., GCN and GAT). However, we observe that serving SGRL models for online applications is challenging due to their irregular request workloads . Specifically, some heavy requests need significantly more computation than regular requests (e.g., 100x), leading to excessively long tail latency in existing systems. As such, we build SG-Serve, which tailors subgraph extraction and model inference (i.e., the two main stages of SGRL models) to handle the irregular request workloads. For subgraph extraction on the CPU, we propose a general API to implement the diverse extraction methods of SGRL models. Beside generality, the API also exposes parallelization opportunities and allows the heavy requests to utilize multiple threads for speedup. To schedule the CPU threads to conduct extraction for concurrent requests, we design a work-stealing policy, which enjoys parallelism while avoiding head-of-line blocking. For model inference on the GPU, we batch the requests according to their workload instead of request count (i.e., as in existing systems) and run two GPU processes to prevent the heavy requests from monopolizing the GPU. Our experiments show that compared with existing systems, SG-Serve can reduce the 99th percentile (P99) latency by over 13x and improve the request throughput by over 33x.
We are excited to announce the fifth issue dedicated to the Principles of Database Systems (PODS) research track of the Proceedings of the ACM on Management of Data (PACMMOD) journal. In its current form, the journal hosts a SIGMOD and a PODS research track. The PODS research track aims to provide a solid scientific basis for methods, techniques, and solutions for the data management challenges that continually arise in our data-driven society. Articles for the PODS track of PACMMOD present principled contributions to modeling, application, system building, and both theoretical and experimental validation in the context of data management. Such articles might be based, among others, on establishing theoretical results, developing new concepts and frameworks that deserve further exploration, providing experimental work that sheds light on the scientific foundations of the discipline, or conducting a rigorous analysis of industry artifacts that are widely used and/or recently developed. At a time when computer science is increasingly data-centric, it is essential to promote an active exchange of tools and techniques between principles of database systems and other communities focused on data management. The PODS research track thus pays special attention to articles that promote the integration of data management techniques within broader computer science.
Index benefit estimation is arguably the most important step in index tuning. Existing index tuners developed for commercial and open-source database systems typically leverage the ''what if'' API for this purpose, which relies on query optimizer's cost estimation module and can be inaccurate due to various reasons such as cardinality estimation errors. Recent work has proposed learning-based index benefit estimators to replace the what-if API. Although learned index benefit estimators show better accuracy, they require large amounts of query execution telemetry as training data, which often contain noisy or conflicting labels that can perplex the trained ML model. There are two types of such noisy labels: epistemic noise, which are caused by the limitations of query plan encodings employed by existing learned estimators and aleatoric noise, which are caused by the inherent dynamicity of the query execution environment due to unpredictable runtime factors such as buffer pool utilization, resource contention, and sometimes adaptive query execution. In this paper, we propose RIB to mitigate the impact of noisy labels and therefore improve the robustness of learned index benefit estimators. RIB introduces two new technologies to address the challenges imposed by epistemic and aleatoric label noises: (1) a context-aware encoder based on bidirectional graph neural network (GNN) and (2) a probabilistic prediction model based on fully parameterized quantile regression (FPQR). Compared to existing work, the GNN-based encoder captures more contextual information about index-optimizable operations, such as structural changes in query plans before and after utilizing indexes, thereby reducing the chance of epistemic noise. Moreover, unlike existing work that uses a point estimate for index benefit estimation, the FPQR-based predictor considers the entire distribution of likely index benefits and provides more robust estimates by aggregating over all quantiles of the distribution, thereby reducing the impact of aleatoric noise. Extensive experiments on top of multiple benchmarks demonstrate that RIB significantly outperforms state-of-the-art index benefit estimators in terms of both estimation accuracy and end-to-end index recommendation quality.
Time series data frequently suffer from data quality problems during collection and transmission, such as small-jump dirty points, which existing cleaning methods often fail to detect. Since existing methods primarily address univariate series, their multivariate extensions often fail to capture complex inter-variable dependencies, significantly limiting their effectiveness. To this end, we propose SHoTClean, a family of four algorithms that bridges hard constraints (i.e., physical limits) and soft constraints (i.e., statistical patterns) within a constrained-optimization framework for effective and efficient multivariate time series cleaning. Specifically, we formulate the cleaning task as minimizing soft-constraint violations while respecting hard-constraint bounds. Then, we propose SHoTClean that introduces: (1) SHoTClean-B for offline batch processing using pruned dynamic programming to achieve global optimality; (2) SHoTClean-S and SHoTClean-P for online streaming scenarios by employing incremental dynamic programming, where SHoTClean-P accelerates SHoTClean-S via CDQ divide-and-conquer and Fenwick tree to attain near-linear complexity; and (3) SHoTClean-C, incorporating causal discovery into soft constraints to capture multivariate dependencies. Extensive experiments across 12 real-world datasets demonstrate that our approaches achieve i) 6.8%--90.0% and 7.8%--82.1% improvements in accuracy (RMSE metric) over 10 state-of-the-art baselines in offline and online settings, respectively; ii) an average two-order-of-magnitude runtime speed-up on large-scale datasets; and iii) superior robustness, with consistent high performance under extreme 80% contamination level and high-dimensional datasets. The code is available at https://github.com/ZJU-DAILY/SHoTClean.
Graph-based Approximate Nearest Neighbor Search (ANNS) is a cornerstone of modern AI systems; however, its practical utility is undermined by a severe long-tail latency problem in which outlier queries jeopardize Service-Level Objectives (SLOs). This paper presents the first systematic study to establish the hubness phenomenon, an intrinsic property of high-dimensional data, as the fundamental cause of this performance instability. Our theoretical analysis reveals that hubness induces topologically skewed proximity graphs, characterized by overly centralized hubs and isolated anti-hubs. This topological imbalance invalidates the greedy traversal heuristic underpinning graph-based search, explaining why queries targeting anti-hubs become performance outliers. We propose a unified framework that deconstructs mainstream ANNS algorithms, reinterpreting their designs as a taxonomy of distinct hubness-mitigation strategies. Extensive experiments on datasets scaling up to 100 million vectors validate this framework, linking an algorithm's mitigation strategy to its effectiveness in balancing graph topology and optimizing outlier performance. Furthermore, our analysis uncovers critical design challenges, specifically the inherent trade-off between suppressing hubs and ensuring the reachability of anti-hubs. Building on these insights, we introduce a hubness-aware pruning optimization that efficiently refines graph topology. This approach yields performance improvements with negligible processing overhead, validating the potential of designing robust graph-based ANNS indexes by integrating hubness information.
Densest subgraph discovery (DSD) is a fundamental research topic in network science and graph databases. As a typical variant of DSD, the anchored densest subgraph (ADS) problem aims to detect a densest subgraph that is anchored around user-defined seed vertices. The ADS problem has been shown to be very useful in some personalized applications, such as community search and recommendation. While the ADS problem is very useful, existing algorithms suffer from weak theoretical guarantees and perform poorly in practice. To address these issues, we first propose a novel approximation algorithm with improved time complexity, achieving the same accuracy as existing methods while requiring significantly fewer iterations. To further enhance practical performance, we introduce a graph reduction technique that localizes the ADS search to a much smaller subgraph, while still providing non-trivial theoretical guarantees. Building on this approximation, we also develop an efficient exact algorithm. We have performed an extensive empirical evaluation of our approaches on 12 real large datasets. The results show that our proposed algorithms are up to four orders of magnitude faster than the state-of-the-art.
Frequency estimation in streaming data often relies on sketches like Count-Min to provide approximate answers with sublinear space. However, Count-Min sketches introduce additive errors that disproportionately impact the unpopular groups, creating fairness concerns. To address these concerns, we introduce Fair-Count-Min, a frequency estimation sketch that guarantees equal expected approximation factors across various groups. We propose a column partitioning approach with group-aware semi-uniform hashing to eliminate collisions between elements from different groups. We provide theoretical guarantees for fairness, analyze its associated cost, and validate our findings through extensive experiments on real-world datasets in comparison with representative state-of-the-art baselines. Our experimental results demonstrate that Fair-Count-Min achieves fairness with generally small additional error while maintaining efficiency comparable to that of the Count-Min sketch.
Learned database components, which deeply integrate machine learning into their design, have been extensively studied in recent years. Given the dynamism of databases, where data and workloads continuously drift, it is crucial for learned database components to remain effective and efficient in the face of data and workload drift. Robustness, therefore, is a key factor in assessing their practical applicability. Although recent works examine learned database components under specific drift, they fail to enable systematic performance evaluations across a broad range of drift or under customized drift as needed. This paper presents NeurBench, a new benchmark suite that supports evaluating learned database components under measurable and controllable data and workload drift. We quantify diverse types of drift by introducing a key concept called the drift factor. Building on this formulation, we propose a drift-aware data and workload generation framework that effectively simulates real-world drift while preserving inherent correlations. Experimental results demonstrate the effectiveness of NeurBench in generating realistic data and workload drift, while providing insights into the performance of representative learned database components under different drift scenarios.
Despite decades of research, the practical adoption of Approximate Query Processing (AQP) remains limited. Sampling-based systems such as VerdictDB often yield slow or inaccurate results for complex analytical queries, as their effectiveness depends heavily on sample quality. Model-based systems such as DeepDB perform well on simple queries but incur high runtime latency for multiway joins and group-by operations, as each group is evaluated as an independent query with instantiated attribute values, preventing global optimization across groups. Both approaches are tied to fixed hardware backends, limiting flexibility across CPU and GPU platforms. This paper presents BayesAQP, a backend-agnostic AQP framework that compiles Bayesian Network (BN) inference into tensor algebra (TA) programs. During training, BayesAQP models per-table statistics using BNs and integrates join histograms for efficient multi-table queries. At runtime, SQL queries are compiled into TA programs in BTL++, a tensor-centric probabilistic programming language, and optimized into tensor contractions for execution on tensor processing frameworks across CPUs and GPUs. Experimental results show that BayesAQP achieves an average 37× speed-up over existing AQP systems on three real-world datasets while maintaining comparable accuracy and significantly improving performance for complex multiway join and group-by queries.
This paper presents an ML-based method for SubIso, the graph pattern matching problem. Given a graph G and a pattern Q, SubIso aims to enumerate all subgraphs of G that are isomorphic to Q. We show that SubIso is EnumP-complete, where EnumP is the enumeration counterpart of NP. We then study revisions VF3M of the classical VF3 algorithm by incorporating an ML oracle M. The model M predicts candidate matches, while VF3M verifies them, thereby reducing costly backtracking. We examine whether VF3M is (a) output-polynomial, i.e., its runtime is polynomial in the sizes of input and output; (b) consistent, i.e., its accuracy reaches 100% with error-free predictions from M; and (c) /3-robust, i.e., it guarantees accuracy of at least /3 even with arbitrarily bad predictions. We establish several results, positive and negative. On the negative side, we show that it is impossible for VF3M to be both output polynomial and /3-robust with a positive constant /3 unless P = fewP, a problem as hard as P = NP. On the positive side, we develop three versions of VF3M that are (a) consistent, and (b) either 1-robust or output polynomial with a high probability. We also train a model M for VF3M. Using real-life and synthetic data, we show that VF3M is up to 15.75 & times;-21 & times; faster than VF3, with F1-score 0.98.
Logic bugs in distributed database management systems silently yield incorrect results and are difficult to detect, yet they threaten correctness in critical deployments. Existing testing techniques focus mainly on system level failures or general DBMS and do not target distributed execution, leaving many distributed logic bugs undiscovered. We present DistSQL, a differential testing framework that compares the results of identical SQL queries executed on the same DBMS configured in centralized and distributed modes. The key insight is that centralized execution is simpler and typically better tested, and thus can serve as a practical reference for distributed execution. DistSQL addresses two difficulties that hinder effective bug finding in distributed settings: it performs distributed diversity oriented database state mutation to expose distribution specific behaviors, and it conducts distributed interaction guided exploration of the execution space using query plan features to prioritize novel behaviors over redundant tests. DistSQL requires no intrusive code instrumentation. The evaluation on five popular open source distributed DBMSs, including TiDB, CockroachDB, YugabyteDB, ClickHouse, and OceanBase, as well as one widely deployed commercial DBMS, demonstrates the efficacy of DistSQL. DistSQL identified 65 previously unknown logic bugs, including 38 specific to distributed execution. Of these, 61 have been confirmed and 50 have been fixed.
Biconnected components (BCCs) are fundamental structures in graph analysis, with applications spanning various domains. To support these applications over continuously evolving data, we study the problem of maintaining BCCs in streaming graphs under the widely used sliding-window model. Existing methods suffer from a severe bottleneck in edge deletion, which dominates their practical running time. To overcome this limitation, we design an index that eliminates the need to maintain BCCs under edge deletions. This index compresses all BCCs across sub-windows ending at the current timestamp and achieves optimal space complexity. When an edge expires and is deleted, the portion of the index corresponding to the sub-window containing this edge can be directly discarded in constant time. For edge insertions, we develop a two-step maintenance framework that progressively transforms the outdated index into the updated version through a sequence of tree-edge rotations. Extensive experiments on real-world datasets demonstrate that our approach considerably outperforms state-of-the-art methods.
Value linking in text-to-SQL systems is the problem of associating values mentioned or implied within a NL query with their database entries. Despite existing efforts, their performance and comparative effectiveness remain unexplored. This work provides the first systematic study focused exclusively on value linking. First, we present a taxonomy that encapsulates the key steps and design choices for value linking. To enable a thorough analysis, we introduce VLD-Bench, a new, challenging benchmark that incorporates 18 categories of lexical and semantic discrepancies between values referenced in the queries and database entries. Our evaluation shows that state-of-the-art text-to-SQL systems suffer up to a 64% drop in execution accuracy due to value linking errors. A fine-grained analysis reveals how design choices impact robustness to linguistic variation and system latency, offering key insights and directions for future research.
Major Database Management Systems (DBMSs) like MySQL field a high volume of bug reports daily. Yet, reproducing database bugs from natural language reports demands substantial developer effort due to manual interpretation and trial-and-error reconstruction. Database bugs are particularly challenging because they involve reconstructing multi-dimensional states that encompass configurations, schemas, data, queries, and validation oracles, while coping with natural language ambiguity. Existing automated bug reproduction techniques, designed for stateless functional testing, cannot address the stateful, multi-dimensional nature of database bugs. We present DBugScribe, the first framework specifically designed to automatically synthesize validated, executable, and structured reproduction scenarios for database bugs directly from users' natural language bug reports. At its core, DBugScribe introduces a domain-specific language (DSL) with formal semantics to represent bug scenarios as composable specifications. A novel hybrid synthesis approach integrates LLM-based information extraction with rule-based validation and self-refinement. Evaluated on 218 confirmed bug reports from eight recent DBMS testing tools covering MySQL, TiDB, and MariaDB, DBugScribe achieves 72.9% reproduction success. In our setup, DBugScribe can synthesize and execute a reproduction scenario within minutes per report. Beyond immediate reproduction, DBugScribe aggregates validated scenarios into a structured knowledge base that enables systematic bug analysis and cross-DBMS bug detection. This work transforms database bug reproduction from a manual, ad-hoc process into an automated and principled discipline, advancing both software engineering practice and database reliability research.
