Vector search plays a crucial role in many real-world applications. In addition to single-vector search, multi-vector search becomes important for multi-modal and multi-feature scenarios today. In a multi-vector database, each row is an item, each column represents a feature of items, and each cell is a high-dimensional vector. In multi-vector databases, the choice of indexes can have a significant impact on performance. Although index tuning for relational databases has been extensively studied, index tuning for multi-vector search remains unclear and challenging. In this paper, we define multi-vector search index tuning and propose a framework to solve it. Specifically, given a multi-vector search workload, we develop algorithms to find indexes that minimize latency and meet storage and recall constraints. Compared to the baseline, our latency achieves 2.1X to 8.3X speedup.

Retrieval plays a fundamental role in recommendation systems, search, and natural language processing (NLP) by efficiently finding relevant items from a large corpus given a query. Dot products have been widely used as the similarity function in such tasks, enabled by Maximum Inner Product Search (MIPS) algorithms for efficient retrieval. However, state-of-the-art retrieval algorithms have migrated to learned similarities. These advanced approaches encompass multiple query embeddings, complex neural networks, direct item ID decoding via beam search, and hybrid solutions. Unfortunately, we lack efficient solutions for retrieval in these state-of-the-art setups. Our work addresses this gap by investigating efficient retrieval techniques with expressive learned similarity functions. We establish Mixture-of-Logits (MoL) as a universal approximator of similarity functions, demonstrate that MoL's expressiveness can be realized empirically to achieve superior performance on diverse retrieval scenarios, and propose techniques to retrieve the approximate top-k results using MoL with tight error bounds. Through extensive experimentation, we show that MoL, enhanced by our proposed mutual information-based load balancing loss, sets new state-of-the-art results across heterogeneous scenarios, including sequential retrieval models in recommendation systems and finetuning language models for question answering; and our approximate top-k algorithms outperform baselines by up to 66× in latency while achieving >.99 recall rate compared to exact algorithms.

There has been considerable interest in leveraging GPUs' computational power and high memory bandwidth for analytical database workloads. However, their limited memory capacity remains a fundamental limitation for databases whose sizes far exceed the GPU memory size. This challenge is exacerbated by the slow PCIe data transfer speed, that creates a bottleneck in overall system performance. In this work, we introduce a hybrid CPU-GPU query processing strategy that leverages the distinct strengths of CPU and GPU to alleviate the data transfer bottleneck. Our approach performs highly efficient data filtering on the CPU, which substantially reduces the volume of data transferred to the GPU via PCIe, and offloads compute-intensive operators such as joins to the GPU for further processing. Our evaluation on the TPC-H benchmark at scale factors up to 1000 (1TB), using a single A100 GPU with 80GB memory, demonstrates that our approach can effectively handle datasets significantly larger than the GPU memory size. Moreover, it substantially outperforms a state-of-the-art CPU-only database system in both performance and cost-effectiveness.

We describe a new benchmark, DSB, for evaluating both workload-driven and traditional database systems on modern decision support workloads. DSB is adapted from the widely-used industrial-standard TPC-DS benchmark. It enhances the TPC-DS benchmark with complex data distribution and challenging yet semantically meaningful query templates. DSB also introduces configurable and dynamic workloads to assess the adaptability of database systems. Since workload-driven and traditional database systems have different performance dimensions, including the additional resources required for tuning and maintaining the systems, we provide guidelines on evaluation methodology and metrics to report. We show a case study on how to evaluate both workload-driven and traditional database systems with the DSB benchmark. The code for the DSB benchmark is open sourced and is available at https://aka.ms/dsb.

Bitvector filtering is an important query processing technique that can significantly reduce the cost of execution, especially for complex decision support queries with multiple joins. Despite its wide application, however, its implication to query optimization is not well understood. In this work, we study how bitvector filters impact query optimization. We show that incorporating bitvector filters into query optimization straightforwardly can increase the plan space complexity by an exponential factor in the number of relations in the query. We analyze the plans with bitvector filters for star and snowflake queries in the plan space of right deep trees without cross products. Surprisingly, with some simplifying assumptions, we prove that, the plan of the minimal cost with bitvector filters can be found from a linear number of plans in the number of relations in the query. This greatly reduces the plan space complexity for such queries from exponential to linear. Motivated by our analysis, we propose an algorithm that accounts for the impact of bitvector filters in query optimization. Our algorithm optimizes the join order for an arbitrary decision support query by choosing from a linear number of candidate plans in the number of relations in the query. We implement our algorithm in a commercial database DBMS-X as a transformation rule. Our evaluation on both industry standard benchmarks and customer workload shows that, compared with DBMS-X, our technique reduces the total CPU execution time by 22%-64% for the workloads, with up to two orders of magnitude reduction in CPU execution time for individual queries.

Recent research has shown promising results by using machine learning (ML) techniques to improve the performance of database systems, e.g., in query optimization or index recommendation. However, in many production deployments, the ML models' performance degrades significantly when the test data diverges from the data used to train these models. In this paper, we address this performance degradation by using B-instances to collect additional data during deployment. We propose an active data collection platform, ADCP, that employs active learning (AL) to gather relevant data cost-effectively. We develop a novel AL technique, Holistic Active Learner (HAL), that robustly combines multiple noisy signals for data gathering in the context of database applications. HAL applies to various ML tasks, budget sizes, cost types, and budgeting interfaces for database applications. We evaluate ADCP on both industry-standard benchmarks and real customer workloads. Our evaluation shows that, compared with other baselines, our technique improves ML models' prediction performance by up to 2x with the same cost budget. In particular, on production workloads, our technique reduces the prediction error of ML models by 75% using about 100 additionally collected queries.

Empirical entropy refers to the information entropy calculated from the empirical distribution of a dataset. It is a widely used aggregation function for knowledge discovery, as well as the foundation of other aggregation functions such as mutual information. However, computing the exact empirical entropy on a large-scale dataset can be expensive. Using a random subsample, we can compute an approximation of the empirical entropy efficiently. We derive probabilistic error bounds for the approximation, where the error bounds reduce in a near square root rate with respect to the subsample size. We further study two applications which can benefit from the error-bounded approximation: feature ranking and filtering based on mutual information. We develop algorithms to progressively subsample the dataset and return correct answers with high probability. The sample complexity of the algorithms is independent of data size. The empirical evaluation of our algorithms on large-scale real-world datasets demonstrates up to three orders of magnitude speedup over exact methods with \errrate\ error.

State-of-the-art index tuners rely on query optimizer's cost estimates to search for the index configuration with the largest estimated execution cost improvement`. Due to well-known limitations in optimizer's estimates, in a significant fraction of cases, an index estimated to improve a query's execution cost, e.g., CPU time, makes that worse when implemented. Such errors are a major impediment for automated indexing in production systems. We observe that comparing the execution cost of two plans of the same query corresponding to different index configurations is a key step during index tuning. Instead of using optimizer's estimates for such comparison, our key insight is that formulating it as a classification task in machine learning results in significantly higher accuracy. We present a study of the design space for this classification problem. We further show how to integrate this classifier into the state-of-the-art index tuners with minimal modifications, i.e., how artificial intelligence (AI) can benefit automated indexing (AI). Our evaluation using industry-standard benchmarks and a large number of real customer workloads demonstrates up to 5x reduction in the errors in identifying the cheaper plan in a pair, which eliminates almost all query execution cost regressions when the model is used in index tuning.

Query performance regression due to the query optimizer selecting a bad query execution plan is a major pain point in production workloads. Commercial DBMSs today can automatically detect and correct such query plan regressions by storing previously-executed plans and reverting to a previous plan which is still valid and has the least execution cost. Such reversion-based plan correction has relatively low risk of plan regression since the decision is based on observed execution costs. However, this approach ignores potentially valuable information of efficient subplans collected from other previously-executed plans. In this paper, we propose a novel technique, Plan Stitch, that automatically and opportunistically combines efficient subplans of previously-executed plans into a valid new plan, which can be cheaper than any individual previously-executed plan. We implement Plan Stitch on top of Microsoft SQL Server. Our experiments on TPC-DS benchmark and three real-world customer workloads show that plans obtained via Plan Stitch can reduce execution cost significantly, with a reduction of up to two orders of magnitude in execution cost when compared to reverting to the cheapest previously-executed plan.

OLTP systems can often improve throughput by batching transactions and processing them as a group. Batching has been used for optimizations such as message packing and group commits; however, there is little research on the benefits of a holistic approach to batching across a transaction's entire life cycle. In this paper, we present a framework to incorporate batching at multiple stages of transaction execution for OLTP systems based on optimistic concurrency control. Storage batching enables reordering of transaction reads and writes at the storage layer, reducing conflicts on the same object. Validator batching enables reordering of transactions before validation, reducing conflicts between transactions. Dependencies between transactions make transaction reordering a non-trivial problem, and we propose several efficient and practical algorithms that can be customized to various transaction precedence policies such as reducing tail latency. We also show how to reorder transactions with a thread-aware policy in multi-threaded OLTP architecture without a centralized validator. In-depth experiments on a research prototype, an opensource OLTP system, and a production OLTP system show that our techniques increase transaction throughput by up to 2.2x and reduce their tail latency by up to 71% compared with the start-of-the-art systems on workloads with high data contention.

Datastores today rely on distribution and replication to achieve improved performance and fault-tolerance. But correctness of many applications depends on strong consistency properties--something that can impose substantial overheads, since it requires coordinating the behavior of multiple nodes. This paper describes a new approach to achieving strong consistency in distributed systems while minimizing communication between nodes. The key insight is to allow the state of the system to be inconsistent during execution, as long as this inconsistency is bounded and does not affect transaction correctness. In contrast to previous work, our approach uses program analysis to extract semantic information about permissible levels of inconsistency and is fully automated. We then employ a novel homeostasis protocol to allow sites to operate independently, without communicating, as long as any inconsistency is governed by appropriate treaties between the nodes. We discuss mechanisms for optimizing treaties based on workload characteristics to minimize communication, as well as a prototype implementation and experiments that demonstrate the benefits of our approach on common transactional benchmarks.

Scaling-out a database system typically requires partitioning the database across multiple servers. If applications do not partition perfectly, then transactions accessing multiple partitions end up being distributed, which has well-known scalability challenges. To address them, we describe a high-performance transaction mechanism that uses optimistic concurrency control on a multi-versioned tree-structured database stored in a shared log. The system scales out by adding servers, without partitioning the database.