With the increasing migration of business operations to the cloud, cloud service providers are facing a growing demand for faster database synchronization across diverse network conditions. Thus, compression methods are predominantly employed over the synchronized database binlog files to reduce the volume of data to be transmitted across the network. However, previous solutions typically rely on using one single compression method. This can result in data compression rates failing to align well with network bandwidth, causing data to wait for compression or transmission, thereby leading to inferior performance. To address the above issues, we propose a fine-grained hybrid adaptive compression system that (1) parses binlog files into multiple fine-grained blocks, and (2) applies a hybrid combination of multiple compression methods to seamlessly align compression speed with the network bandwidth. We have conducted extensive evaluations which demonstrate that, compared to the cutting-edge compression methods like ZSTD, LZ4, and Snappy, our approach can cut down the average latency by 66% and improve the synchronization throughput by 2.45×.

Embedding-based recommendation models (ERMs) require large memory to host huge embedding tables and involve massive data traffic to read the embeddings. As a new interconnect, CXL suits ERMs since it can scale up single-machine memory with performant remote memory devices. However, directly running DRAM-based ERM serving systems on CXL yields poor performance because the bandwidth of CXL is notably lower than DRAM and can be easily saturated, making CXL memory the bottleneck. The non-uniform memory access (NUMA) architecture in modern CXL servers further decreased the system performance. In this paper, we design Carina for ERM serving on heterogeneous memory with CXL by considering such bandwidth asymmetry. In particular, Carina balances the memory access from different memory devices by storing hot embeddings with high access frequencies on DRAM and specifying the placement of embedding tables on the NUMA nodes. Moreover, Carina adopts bandwidth-aware task execution, which decomposes each batch of ERM requests into fine-grained tasks and schedules the tasks to control the real-time utilization of CXL bandwidth to avoid instantaneous saturation. We evaluate Carina under real CXL devices and find that it outperforms a CXL-oblivious baseline by an average of 5.38x and 4.04x in system throughput and request latency, respectively.

Given a set O of objects consisting of n high-dimensional vectors, the problem of approximate nearest neighbor (ANN) search for a query vector q is crucial in many applications where objects are represented as feature vectors in high-dimensional spaces. Each object in O often has attributes like popularity or price, which influence the search. Practically, searching for the nearest neighbor to q might include a range filter specifying the desired attribute values, e.g., within a specific price range. Existing solutions for range filtered ANN search often face trade-offs among excessive storage, poor query performance, and limited support for updates. To address this challenge, we propose RangePQ, a novel indexing scheme that supports efficient range filtered ANN searches and updates, requiring only linear space. Our scheme integrates seamlessly with existing PQ-based index---a widely recognized, scalable index type for ANN searches---to enhance range-filtered ANN queries and update capabilities. Our indexing method, supporting arbitrary range filters, has a space complexity of (O(n log K)), where K is a parameter of the PQ-based index and log K scales with O(log n). To reduce the space cost, we further present a hybrid two-layer structure to reduce space usage to O(n), preserving query efficiency without additional update costs. Experimental results demonstrate that our indexing scheme significantly improves query performance while maintaining competitive update performance and space efficiency.

Distributed coordination services are fundamental components of distributed systems, employing durable replicated state machines (RSMs) to ensure consistency across replicas and prevent data loss, even in the event of all nodes failing. These services typically rely on persistent logs for rapid recovery, as a universally agreed-upon log allows replicas to restore their state by sequentially replaying ordered log entries. However, the requirement for a totally ordered log inherently limits opportunities for parallelism. This paper introduces Fleet, a high-performance durable RSM protocol that combines a hybrid scattered-entry log with an asynchronous ordered log. Our approach integrates synchronous persistence of scattered entries with asynchronous persistence of ordered entries, ensuring both rapid recovery and high levels of parallelism. Additionally, we propose a parallel applying optimization for the etcd database, named pre-apply. Experimental results demonstrate that Fleet significantly outperforms Raft and Scalog in terms of throughput and latency, achieving up to 10× the throughput under specific configurations and scaling effectively across multiple nodes. Additionally, with the pre-apply optimization, Fleet delivers a 10-fold increase in throughput compared to sequential applying on etcd. Although Fleet incurs a 5% overhead in recovery time during leader failure, this delay is tolerable given the rarity of such events.

Real-time data warehouses are essential for modern applications. Extract-Transform-Load (ETL) as a fundamental component of offline data warehouses also provides crucial support within realtime data warehouses. Among various traditional ETL approaches, Lambda and Kappa have emerged as classic real-time data processing solutions due to their freshness and query performance, which best meet business demands. However, both of them often require the integration of external stream processing engines, introducing challenges related to complexity, efficiency, and consistency. ZeroETL has emerged as an approach to address these issues. Nevertheless, existing ZeroETL-based solutions primarily emphasize the implementation of extraction and loading, resulting in limitations in handling transformation. Incremental View Maintenance (IVM) offers an alternative that can enhance ZeroETL. However, existing IVM implementations often focus on query acceleration rather than supporting high-throughput, complex real-time workloads. To address these challenges, we propose Streaming View, an efficient real-time data processing engine integrated within AnalyticDB of Alibaba Cloud. Unlike existing solutions, Streaming View supports high-throughput, complex data processing for realtime streaming ETL workloads. Furthermore, it can be leveraged to optimize ZeroETL-based approaches by enhancing transformation capabilities. We design tailored algorithms and optimizations for diverse syntaxes and high-throughput scenarios, ensuring the system meets complex application needs. By integrating incremental computation into the data warehouse, Streaming View reduces complexity, ensures data consistency, and boosts performance, offering a robust solution for real-world applications. Experiments show Streaming View improves processing performance by up to 7x and 20x over traditional ETL and IVM methods, respectively, and addresses complex scenarios unsolved by existing solutions.

In the era of big data, the landscape of data management and analytics has significantly transformed, presenting diverse challenges for cloud platforms. Modern data warehouses face increasing challenges in handling hybrid transactional and analytical processing (HTAP) workloads efficiently in cloud environments. Traditional shared-nothing architectures provide high-performance query execution but suffer from high storage costs and limited elasticity, while shared-storage approaches improve scalability but often struggle with query efficiency due to increased data movement and indexing overhead. Furthermore, existing execution engines lack optimized support for vectorized processing and real-time analytics, limiting their ability to handle large-scale workloads efficiently. To address these limitations, we introduce AnalyticDB-PG (ADB-PG), a cloud-native, high-performance data warehouse designed for modern analytical workloads. It integrates a unified architecture supporting both Shared-Nothing and Shared-Storage modes, allowing flexible deployment and seamless elasticity. In ADB-PG, we introduce Beam, a hybrid storage engine that efficiently balances row-based and columnar storage for real-time analytics, and Laser, an optimized execution engine leveraging vectorized execution and Just-In-Time compilation to accelerate query processing. The system further incorporates advanced indexing mechanisms, adaptive runtime filtering, and dictionary encoding to enhance performance. Extensive evaluations on TPC-H and TPC-DS benchmarks demonstrate that ADB-PG achieves significant performance improvements while reducing storage and operational costs, making it a compelling solution for modern cloud-based data analytics.

Optimizing distributed protocols has traditionally been a real pain, requiring experts to figure out where improvements can be made, along with rigorous correctness proofs and meticulous implementation. This paper presents a theory to systematize this process. The proposed theory can optimize any existing distributed protocols while preserving all their original quality attributes (e.g., generality, correctness). Crucially, applying the optimizations derived from this theory does not necessitate touching the original implementation --- all you need is just to bolt on a few new message adapters. Case studies demonstrate the effectiveness of this approach. For instance, applying the theory to optimize Spanner's atomic commit protocol results in a mere 4 new message adapters, but improving its latency by up to 51% and peak throughput by up to 56%.

Alibaba relies on its robust database infrastructure to facilitate realtime data access and ensure business continuity despite regional disruptions. To address these operational imperatives, Alibaba developed the Data Transmission Service (DTS), which has become critical for internal applications and public cloud services alike. This paper presents a comprehensive study of the architectural innovations, resource scheduling mechanisms, and performance optimization strategies that have been implemented within DTS to tackle the significant challenges of cross-network, heterogeneous data transmission in a cost-effective manner. We explore the novel Any-to-Any (A2A) architecture, which simplifies the complexity of data paths between diverse databases and mitigates network connectivity issues, thereby significantly reducing development overhead. Additionally, we examine a dynamic network bandwidth scheduling algorithm that effectively maintains Service-Level Objectives (SLOs), complemented by a serverless mechanism that ensures efficient resource utilization. Furthermore, DTS utilizes advanced strategies such as transaction dependency tracking, hot data consolidation, and batching to enhance synchronization performance and efficiency. DTS has distilled the lessons learned from years of serving our customer base and currently supports nearly 1 million public cloud instances annually. Our evaluation results show that DTS can effectively and efficiently handle real-time data transmission in both experimental and production environments.

Two-phase-commit (2PC) has been widely adopted for distributed transaction processing, but it also jeopardizes throughput by introducing two rounds of network communications and two durable log writes to a transaction's critical path. Despite the various proposals that eliminate 2PC such as deterministic database and access localization, 2PC remains the de facto standard since the alternatives often lack generality (e.g., requiring workloads without branches based on query results). In this paper, we present Primo, a distributed transaction protocol that supports a more general set of workloads without 2PC. Primo features write-conflict-free concurrency control that guarantees once a transaction enters the commit phase, no concurrency conflict (e.g., deadlock) would occur when installing the write-set — hence the prepare phase is no longer needed to account for any potential conflict from any partition. In addition, Primo further optimizes the transaction path using asynchronous group commit. With that, the durability delay is also taken off the transaction's critical path. Empirical results on Primo are encouraging – in YCSB and TPC-C, Primo attains 1.42× to 8.25× higher throughput than state-of-the-art general protocols including Sundial and COCO, while having similar latency as COCO which also employs group commit.

Increasingly, cloud database vendors host large-scale geographically distributed clusters to provide cloud database services. When managing the clusters, we observe that it is challenging to simultaneously maximizing the resource allocation ratio and resource availability. This problem becomes more severe in modern cloud database clusters, where resource allocations occur more frequently and on a greater scale. To improve the resource allocation ratio without hurting resource availability, we introduce Eigen, a large-scale cloud-native cluster management system for large-scale databases on the cloud. Based on a resource flow model, we propose a hierarchical resource management system and three resource optimization algorithms that enable end-to-end resource optimization. Furthermore, we demonstrate the system optimization that promotes user experience by reducing scheduling latencies and improving scheduling throughput. Eigen has been launched in a large-scale public-cloud production environment for 30+ months and served more than 30+ regions (100+ available zones) globally. Based on the evaluation of real-world clusters and simulated experiments, Eigen can improve the allocation ratio by over 27% (from 60% to 87.0%) on average, while the ratio of delayed resource provisions is under 0.1%.

Providing ACID transactions under conflicts across globally distributed data is the Everest of transaction processing protocols. Transaction processing in this scenario is particularly costly due to the high latency of cross-continent network links, which inflates concurrency control and data replication overheads. To mitigate the problem, we introduce Ocean Vista - a novel distributed protocol that guarantees strict serializability . We observe that concurrency control and replication address different aspects of resolving the visibility of transactions, and we address both concerns using a multi-version protocol that tracks visibility using version watermarks and arrives at correct visibility decisions using efficient gossip. Gossiping the watermarks enables asynchronous transaction processing and acknowledging transaction visibility in batches in the concurrency control and replication protocols, which improves efficiency under high cross-datacenter network delays. In particular, Ocean Vista can process conflicting transactions in parallel, and supports efficient write-quorum / read-one access using one round trip in the common case. We demonstrate experimentally in a multi-data-center cloud environment that our design outperforms a leading distributed transaction processing engine (TAPIR) more than 10-fold in terms of peak throughput, albeit at the cost of additional latency for gossip. The latency penalty is generally bounded by one wide area network (WAN) round trip time (RTT), and in the best case (i.e., under light load) our system nearly breaks even with TAPIR by committing transactions in around one WAN RTT.

Distributed transactions, which access data items at multiple sites atomically, face well-known scalability challenges. To avoid the high overhead, in prior work Fan et al. proposed Epoch-based Concurrency Control (ECC), which makes transactions visible at epoch boundaries, and presented a system that supports high performance read-only and write-only transactions. However, this idea has a clear difficulty to overcome: the common case of a single transaction that does both reading and writing. This paper proposes ALOHA-DB, a scalable distributed transaction processing system. ALOHA-DB uses a novel paradigm of serializable transaction processing using functors, which conceptually resemble futures in modern programming languages. A functor is a placeholder for the value of a key, which can be computed asynchronously in the future in parallel with other functor computations of the same or other transactions. With multi-versioning in ECC, the functor computations only rely on accessing historical versions, and so the traditional locking mechanism is not needed for concurrency control. Functors elevate ECC to a new level: supporting serializable distributed read-write transactions. This combination of techniques never aborts transactions due to read-write or write-write conflicts, but allows transactions to fail due to logic errors or constraint violations. We used functor-enabled ECC to implement ALOHA-DB and evaluated it using TPC-C and YCSB read-write distributed transactions. Experimental results demonstrate that our system's performance on the TPC-C benchmark is nearly 2 million transactions per second over 20 eight-core virtual machines, which outperforms Calvin, a state-of-the-art transaction processing and replication layer, by one to two orders of magnitude.

There is a trend in recent database research to pursue coordination avoidance and weaker transaction isolation under a long-standing assumption: concurrent serializable transactions under read-write or write-write conflicts require costly synchronization, and thus may incur a steep price in terms of performance. In particular, distributed transactions, which access multiple data items atomically, are considered inherently costly. They require concurrency control for transaction isolation since both read-write and write-write conflicts are possible, and they rely on distributed commitment protocols to ensure atomicity in the presence of failures. This paper presents serializable read-only and write-only distributed transactions as a counterexample to show that concurrent transactions can be processed in parallel with low-overhead despite conflicts.

Distributed shared logs are a powerful building block for distributed systems. By providing fault-tolerant persistence and strong ordering guarantees, applications can use a distributed shared log to reliably communicate a stream of events between processes. This can be used, for example, to replicate application state or to build a reliable publish/subscribe system. The log itself must also replicate data in order to provide availability and fault-tolerance. Key to the design of a distributed shared log is the choice of replication algorithm, which will determine many properties of the system.

Today's online applications depend on fast storage and retrieval of up-to-date data at web scale. To meet this growing demand, the designers of distributed storage systems have devised a rich variety of data replication protocols, offering different trade-offs between consistency, latency, and availability. Understanding the sweet spot, and testing whether a system delivers a particular level of consistency, are challenging problems as consistency itself is difficult to reason about. This demo paper describes an interactive software tool for measuring and visualizing the consistency actually observed by client applications accessing a key-value storage system in real time. The tool can be used to evaluate performance trade-offs in a system with tunable consistency, or to verify the correctness of a storage system that guarantees certain forms of so-called “strong consistency”.

NoSQL storage systems are used extensively by web applications and provide an attractive alternative to conventional databases when the need for scalability outweighs the need for transactions. Several of these systems, notably Amazon's Dynamo and its open-source derivatives, provide quorum-based replication and present the application developer with a choice of multiple client-side "consistency levels" that determine the number of replicas accessed by reads and writes. This setting, in turn, affects both the latency and the consistency observed by the client application. Since using a fixed combination of read and write consistency levels for a given application provides only a limited number of discrete options for tuning the consistency-latency trade-off, we investigate techniques that allow more fine-grained tuning as may be required to support consistency guarantees through service level agreements (SLAs). We consider two such techniques, a novel technique that assigns the consistency level on a peroperation basis by choosing randomly between two options (e.g., weak vs. strong consistency) with a tunable probability, and a known technique that uses weak consistency and injects delays into storage operations artificially. We compare and contrast these two techniques experimentally against each other and against combinations of fixed consistency levels using Apache Cassandra deployed in Amazon's EC2 environment.

A recent paper on benchmarking eventual consistency showed that when a constant workload is applied against Cassandra, the staleness of values returned by read operations exhibits interesting but unexplained variations when plotted against time. In this paper we reproduce this phenomenon and investigate in greater depth the low-level mechanisms that give rise to stale reads. We show that the staleness spikes exhibited by Cassandra are strongly correlated with garbage collection, particularly the "stop-the-world" phase which pauses all application threads in a Java virtual machine. We show experimentally that the staleness spikes can be virtually eliminated by delaying read operations artificially at servers immediately after a garbage collection pause. In our experiments this yields more than a 98% reduction in the number of consistency anomalies that exceed 5ms, and has negligible impact on throughput and latency.