Machine learning applications require timely access to fresh data from primary–backup databases to ensure accurate inference. Existing log replay strategies treat all updates equally and adhere to log-order dependencies, or only prioritize frequently accessed tables, resulting in high latency for prediction queries that usually access a small subset of attributes. Allowing immediate query execution can reduce latency but risks substantial accuracy degradation, as prediction models exhibit varying sensitivity even to minor data staleness. In this paper, we propose AALR, an accuracy-aware log replay strategy that accelerates data visibility for prediction queries while preserving inference accuracy. AALR prioritizes replay at the attribute level, enabling fine-grained replay aligned with query access patterns to avoid unnecessary delay. It leverages a learning-based model to quantify the relationship between data freshness and prediction accuracy, supporting adaptive replay decisions under diverse workloads. Furthermore, AALR introduces an epoch-based two-step replay mechanism, combining column-level parallel classification with row-level latest transaction retention to improve parallelism and resource utilization. Extensive experiments on multiple real-world datasets demonstrate that AALR significantly reduces data visibility latency for prediction queries while maintaining high prediction accuracy, outperforming state-of-the-art log replay strategies.

Large language models (LLMs) advance natural language (NL) interaction with databases by converting queries into SQL. However, users often lack familiarity with database schemas, making it difficult to express precise query requirements. To address this, we propose SQL-QMARS, a multi-agent Text-to-SQL framework designed to interactively clarify user intent. The system evaluates query vagueness using a three-layer metadata structure (theme, table, and field). Based on this evaluation, it dynamically triggers two flows: recommending multi-granular suggestions for vague queries and resolving ambiguities for clear ones. Furthermore, the system supports fusing external data to expand the knowledge source for query suggestions. The demonstration indicates that SQL-QMARS effectively guides users from vague to precise queries, improving the practicality of NL-based database interaction.

Integrating LLMs for data processing enables semantic querying but causes GPU memory bottlenecks and redundant computations. We present SemInfer, an acceleration system for batch semantic processing that treats the KV Cache as a semantic index, offloading pre-computed caches to host storage to eliminate redundancy. To reduce the index size, we propose a pruning strategy based on last-layer aggregated attention to accurately retain critical semantic tokens. Furthermore, we employ a pipeline mechanism to enable the asynchronous overlapping of CPU-GPU transmission and inference computation. This demonstration showcases the complete workflow of SemInfer on the IMDB dataset, achieving up to a 16.3x inference speedup over direct LLM inference and a 90

This study proposes a novel storage engine, SynchroStore, designed to address the inefficiency of update operations in columnar storage systems based on Log-Structured Merge Trees (LSM-Trees) under hybrid workload scenarios. While columnar storage formats demonstrate significant query performance advantages when handling large-scale datasets, traditional columnar storage systems face challenges such as high update complexity and poor real-time performance in data-intensive applications. SynchroStore introduces an incremental row storage mechanism and a fine-grained row-to-column transformation and compaction strategy, effectively balancing data update efficiency and query performance. The storage system employs an in-memory row storage structure to support efficient update operations, and the data is converted to a columnar format after freezing to support high-performance read operations. The core innovations of SynchroStore are reflected in the following aspects:(1) the organic combination of incremental row storage and columnar storage; (2) a fine-grained row-to-column transformation and compaction mechanism; (3) a cost-based scheduling strategy. These innovative features allow SynchroStore to leverage background computational resources for row-to-column transformation and compaction operations, while ensuring query performance is unaffected, thus effectively solving the update performance bottleneck of columnar storage under hybrid workloads. Experimental evaluation results show that, compared to existing columnar storage systems like DuckDB, SynchroStore exhibits significant advantages in update performance under hybrid workloads.

Memory disaggregation has emerged as a promising architecture for improving resource efficiency by decoupling the computing and memory resources. But building efficient range indices in such an architecture faces three critical challenges: (1) coarse-grained concurrency control schemes for coordinating concurrent read/write operations with node splitting incur high contention under the skewed and write-intensive workloads; (2) existing data layouts fail to balance consistency verification and hardware acceleration via SIMD (Single Instruction Multiple Data); and (3) naive caching schemes struggle to adapt to rapidly changing access patterns. To address these challenges, we propose Mariana, a memory-disaggregated skiplist index that integrates three key innovations. First, it uses a fine-grained (i.e., entry-level) latch mechanism combined with dynamic node resizing to minimize the contention and splitting frequency. Second, it employs a tailored data layout for leaf node, which separates keys and values to enable SIMD acceleration while maintaining consistency checks with minimal write overhead. Third, it implements an adaptive caching strategy that tracks node popularity in real-time to optimize network bandwidth utilization during the index traversal. Experimental results show that Mariana achieves 1.7x higher throughput under write-intensive workloads and reduces the P90 latency by 23% under the read-intensive workloads, when comparing to the state-of-the-art indices on disaggregated memory.

With its advanced parallel computing capabilities and high-bandwidth memory, Graphics Processing Unit (GPU) has shown great potential in accelerating structured data analytics pipelines. However, limited GPU memory capacity and frequent data transfer bottlenecks hinder the full exploitation of GPU performance. To address these challenges, this demonstration introduces HeteroDaseX, a CPU-GPU heterogeneity-aware execution engine designed for analytical data processing workloads. Its core innovation lies in constructing a cache for intermediate data that utilizes both CPU memory and GPU memory. When intermediate data required by a processing task is found in the GPU memory cache, subsequent computational steps are seamlessly accelerated on the GPU. We will first delve into the system's design, introducing the management mechanism of the heterogeneous operator cache pool and a heterogeneity-aware cost model. To highlight HeteroDaseX's design considerations for its heterogeneity-aware caching architecture and its superior performance, we have developed a visual interface to demonstrate the system through three scenarios.

Inverted index is a basic tool in many data-intensive applications. Though numerous efforts have been made on efficient inverted index-based query processing, existing schemes do not achieve the expected performance for modern data centers, in which servers are equipped with powerful CPUs and relatively large memory. Through comprehensive measurement studies, we identify the root course is that the data formats for index representation make it unfeasible to design efficient query execution approaches on top of it, which results in poor parallel query support and waste CPU computation. Driven by the findings, we propose to reconcile the in-memory index as columnar structures. To enable this idea, we construct the compact columnar format (i.e., Cocoa) that achieves both desirable space efficiency and maintains the capability for efficient searching support. With Cocoa, we design an efficient query executing scheme that utilizes vectorized batch processing to avoid frequent branch prediction, as well as clause enumeration with pruning to save the overhead of intermediate batch materialization. We build an open-source system VeloSearch to embody our design; experimental results show that VeloSearch achieves ~30× better performance compared with state-of-the-art search libraries such as Lucene and Tantivy.

Distributed in-memory databases are widely adopted to achieve low latency and high bandwidth for data-intensive applications. They support scale-out by sharding and distributing data across multiple nodes. To efficiently adapt to various workloads, distributed in-memory databases must be capable of migrating shards across nodes. In this paper, we demonstrate that state-of-the-art approaches experience significant performance degradation during migration due to service downtime and redundant data transfer. Furthermore, our findings indicate that the presence of service downtime constrains the scalability of migration strategies, while the transfer of redundant data during the snapshot transfer phase limits their adaptability to dynamic workloads. To this end, this paper proposes Aion, a live migration strategy designed for distributed in-memory databases. Aion eliminates any potential service downtime by immediately switching transaction routing to the destination node. To ensure data consistency between the source and destination nodes, as well as serializable execution during migration, Aion proposes the mutual validation phase. Moreover, Aion introduces an analysis phase before the snapshot transfer phase to identify dynamically changing hotspots in workloads. The analysis phase identifies and transfers tuples and versions accessed less frequently to the destination node, reducing the amount of data transferred. Aion is implemented on a distributed in-memory database and evaluated using various OLTP workloads. The results demonstrate that Aion can fundamentally eliminate service downtime, adapt effectively to various workloads and exhibit robust scalability. Compared to state-of-the-art approaches, Aion achieves up to 2.25x–6.57x higher throughput during migration and shortens the migration duration by 53.7–68.2

Previous learning-based parameter query optimization (PQO) methods excel in static workloads by precisely selecting optimal plans in a cache with a fixed set of representative plans. However, these methods struggle in dynamic workloads because they cannot predict over dynamically changed plan caches. These queries that fall outside the previously observed query parameter distribution have the risk of reusing suboptimal plans. Unlike traditional PQO methods that learn mappings from parametric query parameters to a fixed set of execution plans, our adaptive PQO framework ( APQO ) takes both query parameters and the plans themselves as model inputs, thereby handling variable numbers of plans in dynamic workloads. By embedding plan representations through representation learning, we pre-train a foundation model offline, enabling APQO to acquire a generalizable plan performance prediction model. Leveraging the foundation model's predictive capability along with a hybrid data augmentation strategy, we train an online calibration model with minimal training data for distribution-shifted new queries, rapidly adapting knowledge for reusing new plans. APQO is natively designed to handle the characteristics of dynamic workloads. Experimental results show that APQO outperforms existing PQO methods in dynamic workloads, achieving a higher cache hit ratio and significantly reducing query latency.

Single-master multi-slave is the mainstream architecture of cloud-native databases.In the cluster,slave nodes can share the read-only requests of the master node,while write requests are handled by the master node.Based on this,to further meet the demands of large-scale transaction expansion,some cloud databases attempt to implement multi-write transaction expansion.One possible approach to multi-write expansion is to introduce shared cache among computing nodes to support cross-node data access.For shared-cache database systems,the overhead of cross-node remote access is significantly higher than that of local access.Therefore,the design of cache protocol is a crucial factor that affects system performance and scalability.This study proposes two innovative improvements to the coherence protocol and implements PG-RAC,a shared-cache database,which supports multi-write transactions based on PostgreSQL.On one hand,PG-RAC proposes a new distributed chained routing strategy,which disperses routing information among computing nodes.Compared to the routing strategy that utilizes single-node directory management,it reduces the average transaction latency by approximately 20%.On the other hand,this study also enhances the duplicate page invalidation mechanism by separating invalidation operations from the transaction path,reducing the latency of the critical path in the transaction.Based on this,PG-RAC takes advantage of the characteristics of multi-version concurrency control(MVCC)and further proposes to delay the invalidation point of duplicate pages,which effectively improves cache utilization.TPC-C experimental results show that for a cluster with 4 compute nodes,the throughput is nearly 2 times that of PostgreSQL and 1.5 times that of the distributed database Citus.

In a geo-distributed database, data shards and their respective replicas are deployed in distinct datacenters across multiple regions, enabling regional-level disaster recovery and the ability to serve global users locally. However, transaction processing in geo-distributed databases requires multiple cross-region communications, especially during the commit phase, which can significantly impact system performance. To optimize the performance of geo-distributed transactions, we propose Decentralized Two-phase Commit (D2PC), a new transaction commit protocol aiming to minimize the negative impact of cross-region communication. In D2PC, we employ multiple co-coordinators that perform commit coordination in parallel. Each co-coordinator is responsible for collecting 2PC votes and making a PreCommit decision in its local region. This approach allows for the concurrent invocation of multiple cross-region network round trips, and each region can end its concurrency control locally before replication is complete, thus significantly reducing the chances of blocking and enhancing system concurrency. Moreover, we propose the bypass leader replication reply method, leveraging decentralized co-coordinators to bypass the leader for message transmission, thereby reducing the commit latency. Experimental results have demonstrated that D2PC can reduce commit latency by 43% and improve throughput by up to 2.43 × compared to the geo-distributed transaction processing methods based on 2PC.

Functionalities of a database system are co-designed and jointly maintain the database performance. Each function-ality usually has its own metrics to evaluate its state. Previous knobs tuning methods regard the database system as a black box and aim to automatically find the optimal configurations by collecting and observing the overall performance data (e.g., transaction throughput per second) under various configuration knobs. However, if a functionality is not running in the tuning phase, its knobs irrelevant to performance changes can also be tuned by existing tools and potential risks would be introduced. To resolve this problem, we design a database knob tuning framework to support functionality-aware knobs tuning. It uses multitask learning to take the database overall performance as the objective of main learning task, and each function module as a separate learning task. This framework enhances the tuning results through learning the relationships between different tasks, and avoids adjusting irrelevant knobs by perceiving the status of functionalities. We validate its generalizability on OceanBase and PostgreSQL. Experimental results show that better performances were achieved on the overall performance and the metrics of various functionalities.

This paper introduces a composable architecture for cloud database systems, which is adaptable to heterogeneous workloads of real-world applications. Based on this architecture, a transactional database can be decomposed into a collection of building blocks, each of which can independently choose its implementation and its way to scale on the heterogeneous resources in the cloud. These building blocks include A concurrency control block enforcing serializable schedules of transactions, A log replication block ensuring the persistence of transactions and high availability, and A data storage block serving data accesses. Such an architecture would make implementing a cloud database more agile and significantly improve its flexibility and application range. Some preliminary experiments were conducted to demonstrate the practicality of the proposed composable architecture.

Cache replacement policies are essential for maximizing application performance. Policies such as LRFU, which incorporate both recency and frequency, have shown efficacy in improving hit rates in many studies. In this paper, we theoretically investigated how parameters impact hit rates in LRFU and discovered two distinct features named unimodality and correlation. Drawing on our understanding, we formulated Hill-Cache. Hill-Cache provides a holistic approach to cache. It incorporates recency and frequency and employs a hill-climbing algorithm for adaptability. Additionally, it improves churn resistance through quick demotion, the approach that the most recent research suggests. By overcoming the limitations of LRFU, including high maintenance overhead and dependency on parameters, Hill-Cache distinguishes itself as a new cache method. Hill-Cache is well-suited for situations where there is an efficiency gap between performance devices and capacity devices, including data caching applications and database systems. Our evaluations, employing 36 real-world traces and various sophisticated policies across diverse cache sizes, demonstrated the superior performance of Hill-Cache. It reduces the average miss rate by 11.97% compared to LRU and outperforms other advanced cache policies such as ARC, LIRS, DLIRS, CACHEUS, and S3FIFO. We incorporated Hill-Cache into Memcached and RocksDB, significantly improving performance metrics such as throughput and latency.

Learning-based Parametric Query Optimization (PQO) methods excel in static workloads with precise cached plan selection but struggle with dynamic workloads. When facing a parametric query outside query parameter distribution in its training workload, suboptimal plan could be selected and this plan would be unsafely reused. The root cause of this limitation is the plan selection model cannot adapt to distribution drifts in query parameters. In order to extend learning-based Parametric Query Optimization for safely reusing cached plans in dynamic workloads, we introduce a novel approach to predict and avoid reusing a suboptimal plan, referred to as SPQO. As each cached plan has specific reuse decision boundary, each cached plan is assigned to an independent binary classifier. In the offline phase, we employ an under-sampling algorithm integrated with Tomek Links technique to effectively train these classifiers under class imbalance setting. During the online phase, we implement hybrid adjustment strategies based on incremental learning, continuously training these classifiers with each prediction and query feedback. Our experiments show SPQO can reduce the 95th percentile relative query latency by 10× in static and 10^3× in dynamic workloads, and achieve better cache hit rates on various workloads.

The CPU overhead of the LSM-tree becomes increasingly significant when high-speed storage devices are utilized. In this article, we propose SuccinctKV, a key-value store based on LSM-tree that is optimized to improve CPU efficiency in mixed workload scenarios. To achieve this, SuccinctKV reduces the CPU overhead of compaction by writing scan-sorted data directly to the storage device. SuccinctKV also redesigns the merge-sort operation of the LSM-tree, enhancing CPU locality and reducing the unnecessary CPU overhead of cache accesses and I/O system calls. Additionally, SuccinctKV introduces a scheduler to resolve potential bursty I/O contention by autonomously initiating I/O requests at appropriate times and quickly relieving I/O pressure by terminating background I/O requests. We implement SuccinctKV on RocksDB and conduct extensive experiments to evaluate our proposed methods. The experimental results demonstrate that, compared with RocksDB, SuccinctKV achieves a maximum improvement of 2.6x in scan performance and reduces CPU overhead of compaction by up to 89% under mixed workloads.

特征管理是搭建人工智能数据管道中的重要一环.特征存储要求在模型训练和推理阶段提供有效版本的特征推送服务.为响应这一需求,特征存储需要为特征实时更新和版本管理提供保证,以协同上游的特征摄取,为模型服务系统提供数据动力.在人工智能辅助决策的在线预测任务中,为了提供更好的用户体验,模型服务系统需要实时响应决策请求,实时特征检索面临更低延迟的挑战.聚焦这一挑战,开发基于内存的多版本在线特征存储FeaDB.使用时间序列建模特征,并提供特征版本管理语义,满足特征从生产到消费的版本管理需求;采用追加写方式保证实时特征加载性能,设计基于版本的索引减少读延迟;为进一步减小特征消费延迟,提出版本快照机制,实验证明采用快照读机制增加了特征集版本的检索效率.

随着物联网的发展,大量传感器设备被接入网络,这些设备所产生的数据的异常检测关系到系统服务的稳定性.时序数据库是专门为时序数据优化的数据库系统.作为监控系统的重要环节,时序数据库担负着时序数据的管理和查询任务.但目前时序数据库在处理多个数据源数据的查询上存在着延迟高、没有充分利用系统计算资源的缺点.针对上述缺点,基于InfluxDB,重新设计了时序数据库的查询执行模型,提出了InfluxDB-PP(parallel processing),很好地解决了上述问题.实验结果表明,InfluxDB-PP在实时数据异常查询场景下,查询时延相较于InfluxDB降低了约85.7％.

In this paper, we present PM-Blade, an LSM-tree structured storage augmented with persistent memory (or non-volatile memory). PM-Blade utilizes persistent memory to optimize read performance and reduce write amplification, which are essential to Meituan’s online retail applications. Distinguished from existing designs, PM-Blade leverages persistent memory to drastically increase the capacity of the level-0 layer of LSM-tree. An enlarged level-0 layer allows a large amount of hot or warm data to be retained in persistent memory, enabling high read performance. At the same time, it works as a large write buffer that absorbs write amplification. To make the best of the design, we devised an internal compaction method and used a cost-based compaction strategy to maximize the utility of the level-0 layer. We implemented the compaction method using coroutines to improve its efficiency and resource utilization. We evaluated PM-Blade through extensive experiments, in which PM-Blade outperformed several open-source alternatives on standard benchmarks and a real-world workload of Meituan.

云原生数据库的崛起,使得共享缓存架构再度受到重视.虽然共享缓存架构能够有效地解决多个读写节点间的缓存一致性问题,但其他问题仍然存在,例如持久化速度缓慢、维护缓存目录的延迟高、时间戳瓶颈等.针对上述问题,提出了一种基于共享缓存架构,并结合新型硬件——持久化内存的解决方案,从而构建了一个包括内存层、持久化内存层、存储层的三层共享架构数据库——TampoDB.基于此架构,重新设计了事务的执行流程,并对时间戳和目录进行了优化,以解决上述问题.实验结果表明,TampoDB有效地提高了事务的持久化速度.