Distributed key-value stores are widely adopted to support elastic big data applications, leveraging purpose-built consensus algorithms like Raft to ensure data consistency. However, through systematic analysis, we reveal a critical performance issue in such consistent stores, i.e., overlapping persistence operations between consensus protocols and underlying storage engines result in significant I/O overhead. To address this issue, we present Nezha, a prototype distributed storage system that innovatively integrates key-value separation with Raft to provide scalable throughput in a strong consistency guarantee. Nezha redesigns the persistence strategy at the operation level and incorporates leveled garbage collection, significantly improving read and write performance while preserving Raft's safety properties. Experimental results demonstrate that, on average, Nezha achieves throughput improvements of 460.2%, 12.5%, and 72.6% for put, get, and scan operations, respectively.

Logic bugs can cause DBMSs to silently produce incorrect results for a given query, posing significant threats to software reliability and remaining challenging to detect. Join is a fundamental operation in DBMSs, enabling the combination of data from multiple tables; however, due to its complexity, it is also susceptible to logic bugs. Existing works detect logic bugs in join optimizations by altering query hints and system variables to alter the optimizer's choice of execution plans. However, these approaches struggle to detect logic bugs when query hints or system variables fail to influence the optimizer's behavior, or when the logic bugs reside in join implementation code that is unrelated to optimization. In this paper, we present S et R elation S ynthesis (SRS), a black-box testing approach that detects logic bugs of join implementation in DBMSs by leveraging set relations among different join operations. SRS applies transformations to the original join queries, including modifications to join types, join orders, and join conditions, while ensuring that the outputs of both the original and transformed queries preserve the expected set relations. Violations of these set relations indicate potential logic bugs. We realized SRS and evaluated it on five widely-used and extensively-tested DBMSs: MySQL, MariaDB, TiDB, PostgreSQL, and DuckDB. SRS uncovered 33 previously unknown and unique bugs, all of which have been confirmed, with 12 already fixed. Among these, 33 are logic bugs, demonstrating SRS's effectiveness and practicality in detecting logic bugs in the implementation of join operations within DBMSs.

Graph Neural Networks (GNNs) have been increasingly adopted for graph analysis in web applications such as social networks. Yet, efficient GNN serving remains a critical challenge due to high workload fluctuations and intricate GNN operations. Serverless computing, thanks to its flexibility and agility, offers on-demand serving of GNN inference requests. Alas, the request-centric serverless model is still too coarse-grained to avoid resource waste.

Abstract Re-ranking is to refine the candidate ranking list of recommended items, such that the re-ranked list attracts users to purchase or click more items than the candidate one without re-ranking. Items in the candidate list are often ranked by their relevance to users’ interests. It is thus important to exploit the mutual influence between items in the re-ranking process. Existing re-ranking models focus on only the pairwise influence between two items, and have limited capability to exploit the local mutual influence in a group of items. Users often show successive interests on a group of relevant items, e.g., mobile phone, phone covers, wireless headset, namely scene. We propose a novel re-ranking model that jointly exploits the local mutual influence in scenes and the global mutual influence between different scenes. Scene representations are learned by GNN and multi-head attention, where GNN aims to learn local mutual influence while multi-head attention is to learn global mutual influence. To study the interaction between users and scenes, matrix factorization on users is utilized to obtain the user preference, which can be further applied to scenes to compute the scene scores. The final re-ranking list is generated by sorting the predicted scores of all scenes. To further mine user history information and item related user information, we also develop the extension pretraining module which relies on mask mechanism to support users and items high-quality embedding generation. We conduct a comprehensive evaluation on several real-world datasets. The experimental results demonstrate that our model substantially outperforms existing approaches.

Sharing data and computation among concurrent queries has been an active research topic in database systems. While work in this area developed algorithms and systems that are shown to be effective, there is a lack of logical foundation for query processing and optimization. In this paper, we present PsiDB, a system model for processing a large number of database queries in a batch. The key idea is to generate a single query expression that returns a global relation containing all the data needed for individual queries. For that, we propose the use of a type of relational operators called ψ-operators in combining the individual queries into the global expression. We tackle the algebraic optimization problem in PsiDB by developing equivalence rules to transform concurrent queries with the purpose of revealing query optimization opportunities. Centering around the ψ-operator, our rules not only cover many optimization techniques adopted in existing batch processing systems, but also revealed new optimization opportunities. Experiments conducted on an early prototype of PsiDB show a performance improvement of up to 36X over a mainstream commercial DBMS.

The top-k dominating (TKD) query on skyline groups returns k skyline groups that dominate the maximum number of points in a given data set. The TKD query combines the advantages of skyline groups and top-k dominating queries, thus has been frequently used in decision making, recommendation systems, and quantitative economics. Traditional skylines are inadequate to answer queries from both individual and groups of points. The group size could be too large to be processed in a reasonable time as a single operator (i.e., the skyline group operator). In this paper, we address the performance problem of grouping for TKD queries in skyline database. We formulate the problem of grouping, define the group operator in skyline, and propose several efficient algorithms to find top-k skyline groups. Thus, we provide a systematic study of TKD queries on skyline groups and validate our algorithms with extensive empirical results on synthetic and realworld data.

While work in Techniques based on sharing data and computation among queries developed algorithms and systems that are shown to be effective, there is a lack of logical foundation for query processing and optimization. In this paper, we present PsiDB, a system model for processing a large number of database queries in a batch. The key idea is to generate a single query expression that returns a global relation containing all the data needed for individual queries. For that, we propose the use of a type of relational operators called ψ-operators in combining the individual queries into the global expression. We tackle the algebraic optimization problem in PsiDB by developing equivalence rules to transform concurrent queries with the purpose of revealing query optimization opportunities. Experiments conducted on an early prototype of PsiDB show a major performance improvement over a mainstream commercial DBMS.

Energy conservation is a growing important issue in designing modern database management system (DBMS). This requires a deep thinking about the tradeoffs between energy and performance. Despite the significant amount of efforts at the hardware level to make the major components consume less energy, we argue for a revisit of the DBMS query processing mechanism to identify and harvest the potential of energy saving. However, the state-of-art architecture of DBMS does not take energy usage into consideration in its design. A major challenge in developing an energy-aware DBMS is to design and implement a cost-based query optimizer that evaluates query plans by both performance and energy costs. By following such a strategy, our previous work revealed the fact that energy-efficient query plans do not necessarily have the shortest processing time. This demo proposal introduces PET -- an energy-aware query optimization framework that is built as a part of the PostgreSQL kernel. PET, via its power cost estimation module and plan evaluation model, enables the database system to run under a DBA-specified energy/performance tradeoff level. PET contains a power cost estimator that can accurately estimate the power cost of query plans at compile time, and a query evaluation engine that the DBA could configure key PET parameters towards the desired tradeoff. The software to be demonstrated will also include workload engine for producing large quantities of queries and data sets. Our demonstration will show how PET functions via a comprehensive set of views from its graphical user interface named PET Viewer . Through such interfaces, a user can achieve a good understanding of the energy-related query optimization and cost-based plan generation. Users are also allowed to interact with PET to experience the different energy/performance tradeoffs by changing PET and workload parameters at query runtime.

With the total energy consumption of computing systems increasing in a steep rate, much attention has been paid to the design of energy-efficient computing systems and applications. So far, database system design has focused on improving performance of query processing. The objective of this study is to experimentally explore the potential of power conservation in relational database management systems. We hypothesize that, by modifying the query optimizer in a DBMS to take the power cost of query plans into consideration, we will be able to reduce the power usage of database servers and control the tradeoffs between power consumption and system performance. We also identify the sources of such savings by investigating the resource consumption features during query processing in DBMSs. To that end, we provide an in-depth anatomy and qualitatively analyze the power profile of typical queries in the TPC benchmarks. We perform extensive experiments on a physical testbed based on the PostgreSQL system using workloads generated from the TPC benchmarks. Our hypothesis is supported by such experimental results: power savings in the range of 11% - 22% can be achieved by equipping the DBMS with a query optimizer that selects query plans based on both estimated processing time and power requirements.1