Computer Science

Configuring databases for efficient querying is a complex task, often carried out by a database administrator. Solving the problem of building indexes that truly optimize database access requires a substantial amount of database and domain knowledge, the lack of which often results in wasted space and memory for irrelevant indexes, possibly jeopardizing database performance for querying and certainly degrading performance for updating. We develop an architecture to solve the problem of automatically indexing a database by using reinforcement learning to optimize queries by indexing data throughout the lifetime of a database. In our experimental evaluation, our architecture shows superior performance compared to related work on reinforcement learning and genetic algorithms, maintaining near-optimal index configurations and efficiently scaling to large databases.

Distributed Systems Infrastructure (DSI) is MongoDB's framework for running fully automated system performance tests in our Continuous Integration (CI) environment. To run in CI it needs to automate everything end-to-end: provisioning and deploying multi-node clusters, executing tests, tuning the system for repeatable results, and collecting and analyzing the results. Today DSI is MongoDB's most used and most useful performance testing tool. It runs almost 200 different benchmarks in daily CI, and we also use it for manual performance investigations. As we can alert the responsible engineer in a timely fashion, all but one of the major regressions were fixed before the 4.2.0 release. We are also able to catch net new improvements, of which DSI caught 17. We open sourced DSI in March 2020.

Companies and individuals produce numerous tabular data. The objective of this position paper is to draw up the challenges posed by the automatic integration of data in the form of tables so that they can be cross-analyzed. We provide a first automatic solution for the integration of such tabular data to allow On-Line Analysis Processing. To fulfil this task, features of tabular data should be analyzed and the challenge of automatic multidimensional schema generation should be addressed. Hence, we propose a typology of tabular data and discuss our idea of an automatic solution.

In the use of database systems, the design of the storage engine and data model directly affects the performance of the database when performing queries. Therefore, the users of the database need to select the storage engine and design data model according to the workload encountered. However, in a hybrid workload, the query set of the database is dynamically changing, and the design of its optimal storage structure is also changing. Motivated by this, we propose an automatic storage structure selection system based on learning cost, which is used to dynamically select the optimal storage structure of the database under hybrid workloads. In the system, we introduce a machine learning method to build a cost model for the storage engine, and a column-oriented data layout generation algorithm. Experimental results show that the proposed system can choose the optimal combination of storage engine and data model according to the current workload, which greatly improves the performance of the default storage structure. And the system is designed to be compatible with different storage engines for easy use in practical applications.

Virtually all of today's Big Data systems are passive in nature, responding to queries posted by their users. Instead, we are working to shift Big Data platforms from passive to active. In our view, a Big Active Data (BAD) system should continuously and reliably capture Big Data while enabling timely and automatic delivery of relevant information to a large pool of interested users, as well as supporting retrospective analyses of historical information. While various scalable streaming query engines have been created, their active behavior is limited to a (relatively) small window of the incoming data. To this end we have created a BAD platform that combines ideas and capabilities from both Big Data and Active Data (e.g., Publish/Subscribe, Streaming Engines). It supports complex subscriptions that consider not only newly arrived items but also their relationships to past, stored data. Further, it can provide actionable notifications by enriching the subscription results with other useful data. Our platform extends an existing open-source Big Data Management System, Apache AsterixDB, with an active toolkit. The toolkit contains features to rapidly ingest semistructured data, share execution pipelines among users, manage scaled user data subscriptions, and actively monitor the state of the data to produce individualized information for each user. This paper describes the features and design of our current BAD data platform and demonstrates its ability to scale without sacrificing query capabilities or result individualization.

Query optimization remains one of the most challenging problems in data management systems. Recent efforts to apply machine learning techniques to query optimization challenges have been promising, but have shown few practical gains due to substantive training overhead, inability to adapt to changes, and poor tail performance. Motivated by these difficulties and drawing upon a long history of research in multi-armed bandits, we introduce Bao (the BAndit Optimizer). Bao takes advantage of the wisdom built into existing query optimizers by providing per-query optimization hints. Bao combines modern tree convolutional neural networks with Thompson sampling, a decades-old and well-studied reinforcement learning algorithm. As a result, Bao automatically learns from its mistakes and adapts to changes in query workloads, data, and schema. Experimentally, we demonstrate that Bao can quickly (an order of magnitude faster than previous approaches) learn strategies that improve end-to-end query execution performance, including tail latency. In cloud environments, we show that Bao can offer both reduced costs and better performance compared with a sophisticated commercial system.

Given a stream of food orders and available delivery vehicles, how should orders be assigned to vehicles so that the delivery time is minimized? Several decisions have to be made: (1) assignment of orders to vehicles, (2) grouping orders into batches to cope with limited vehicle availability, and (3) adapting to dynamic positions of delivery vehicles. We show that the minimization problem is not only NP-hard but inapproximable in polynomial time. To mitigate this computational bottleneck, we develop an algorithm called FoodMatch, which maps the vehicle assignment problem to that of minimum weight perfect matching on a bipartite graph. To further reduce the quadratic construction cost of the bipartite graph, we deploy best-first search to only compute a subgraph that is highly likely to contain the minimum matching. The solution quality is further enhanced by reducing batching to a graph clustering problem and anticipating dynamic positions of vehicles through angular distance. Extensive experiments on food-delivery data from large metropolitan cities establish that FoodMatch is substantially better than baseline strategies on a number of metrics, while being efficient enough to handle real-world workloads.

An increasing number of entities are described by interlinked data rather than documents on the Web. Entity Resolution (ER) aims to identify descriptions of the same real-world entity within one or across knowledge bases in the Web of data. To reduce the required number of pairwise comparisons among descriptions, ER methods typically perform a pre-processing step, called \emph{blocking}, which places similar entity descriptions into blocks and thus only compare descriptions within the same block. We experimentally evaluate several blocking methods proposed for the Web of data using real datasets, whose characteristics significantly impact their effectiveness and efficiency. The proposed experimental evaluation framework allows us to better understand the characteristics of the missed matching entity descriptions and contrast them with ground truth obtained from different kinds of relatedness links.

The growing interest in making use of Knowledge Graphs for developing explainable artificial intelligence, there is an increasing need for a comparable and repeatable comparison of the performance of Knowledge Graph-based systems. History in computer science has shown that a main driver to scientific advances, and in fact a core element of the scientific method as a whole, is the provision of benchmarks to make progress measurable. This paper gives an overview of benchmarks used to evaluate systems that process Knowledge Graphs.

Recent advancements in learned index structures propose replacing existing index structures, like B-Trees, with approximate learned models. In this work, we present a unified benchmark that compares well-tuned implementations of three learned index structures against several state-of-the-art "traditional" baselines. Using four real-world datasets, we demonstrate that learned index structures can indeed outperform non-learned indexes in read-only in-memory workloads over a dense array. We also investigate the impact of caching, pipelining, dataset size, and key size. We study the performance profile of learned index structures, and build an explanation for why learned models achieve such good performance. Finally, we investigate other important properties of learned index structures, such as their performance in multi-threaded systems and their build times.