AWS Glue is Amazon's serverless data integration cloud service that makes it simple and cost effective to extract, clean, enrich, load, and organize data. Originally launched in August 2017, AWS Glue began as an extract-transform-load (ETL) service designed to relieve developers and data engineers of the undifferentiated heavy lifting needed to load databases, data warehouses, and build data lakes on Amazon S3. Since then, it has evolved to serve a larger audience including ETL specialists and data scientists, and includes a broader suite of data integration capabilities. Today, hundreds of thousands of customers use AWS Glue every month. In this paper, we describe the use cases and challenges cloud customers face in preparing data for analytics and the tenets we chose to drive Glue's design. We chose early on to focus on ease-of-use, scale, and extensibility. At its core, Glue offers serverless Apache Spark and Python engines backed by a purpose-built resource manager for fast startup and auto-scaling. In Spark, it offers a new data structure --- DynamicFrames --- for manipulating messy schema-free semi-structured data such as event logs, a variety of transformations and tooling to simplify data preparation, and a new shuffle plugin to offload to cloud storage. It also includes a Hivemetastore compatible Data Catalog with Glue crawlers to build and manage metadata, e.g. for data lakes on Amazon S3. Finally, Glue Studio is its visual interface for authoring Spark and Python-based ETL jobs. We describe the innovations that differentiate AWS Glue and drive its popularity and how it has evolved over the years.

Recent architectural trends---cheap, fast solid-state storage, inexpensive DRAM, and multi-core CPUs---provide an opportunity to rethink the interface between applications and persistent storage. To leverage these advances, we propose a new system architecture called Hathi that provides an in-memory transactional heap made persistent using high-speed flash drives. With Hathi, programmers can make consistent concurrent updates to in-memory data structures that survive system failures.

Energy is a growing component of the operational cost for many "big data" deployments, and hence has become increasingly important for practitioners of large-scale data analysis who require scale-out clusters or parallel DBMS appliances. Although a number of recent studies have investigated the energy efficiency of DBMSs, none of these studies have looked at the architectural design space of energy-efficient parallel DBMS clusters. There are many challenges to increasing the energy efficiency of a DBMS cluster, including dealing with the inherent scaling inefficiency of parallel data processing, and choosing the appropriate energy-efficient hardware. In this paper, we experimentally examine and analyze a number of key parameters related to these challenges for designing energy-efficient database clusters. We explore the cluster design space using empirical results and propose a model that considers the key bottlenecks to energy efficiency in a parallel DBMS. This paper represents a key first step in designing energy-efficient database clusters, which is increasingly important given the trend toward parallel database appliances.

Data management systems have traditionally been designed to support either long-running analytics queries or short-lived transactions, but an increasing number of applications need both. For example, online games, socio-mobile apps, and e-commerce sites need to not only maintain operational state, but also analyze that data quickly to make predictions and recommendations that improve user experience. In this paper, we present Minuet, a distributed, main-memory B-tree that supports both transactions and copy-on-write snapshots for in-situ analytics. Minuet uses main-memory storage to enable low-latency transactional operations as well as analytics queries without compromising transaction performance. In addition to supporting read-only analytics queries on snapshots, Minuet supports writable clones, so that users can create branching versions of the data. This feature can be quite useful, e.g. to support complex "what-if" analysis or to facilitate wide-area replication. Our experiments show that Minuet outperforms a commercial main-memory database in many ways. It scales to hundreds of cores and TBs of memory, and can process hundreds of thousands of B-tree operations per second while executing long-running scans.

Motivated by the increasing popularity of eventually consistent key-value stores as a commercial service, we address two important problems related to the consistency properties in a history of operations on a read/write register (i.e., the start time, finish time, argument, and response of every operation). First, we consider how to detect a consistency violation as soon as one happens. To this end, we formulate a specification for online verification algorithms, and we present such algorithms for several well-known consistency properties. Second, we consider how to quantify the severity of the violations, if a history is found to contain consistency violations. We investigate two quantities: one is the staleness of the reads, and the other is the commonality of violations. For staleness, we further consider time-based staleness and operation-count-based staleness. We present efficient algorithms that compute these quantities. We believe that addressing these problems helps both key-value store providers and users adopt data consistency as an important aspect of key-value store offerings.

Cloud computing demands cheap, always-on, and reliable storage. We describe Pahoehoe, a key-value cloud storage system we designed to store large objects cost-effectively with high availability. Pahoehoe stores objects across multiple data centers and provides eventual consistency so to be available during network partitions. Pahoehoe uses erasure codes to store objects with high reliability at low cost. Its use of erasure codes distinguishes Pahoehoe from other cloud storage systems, and presents a challenge for efficiently providing eventual consistency. We describe Pahoehoe's put, get, and convergence protocols-convergence being the decentralized protocol that ensures eventual consistency. We use simulated executions of Pahoehoe to evaluate the efficiency of convergence, in terms of message count and message bytes sent, for failure-free and expected failure scenarios (e.g., partitions and server unavailability). We describe and evaluate optimizations to the naïve convergence protocol that reduce the cost of convergence in all scenarios.

Rising energy costs in large data centers are driving an agenda for energy-efficient computing. In this paper, we focus on the role of database software in affecting, and, ultimately, improving the energy efficiency of a server. We first characterize the power-use profiles of database operators under different configuration parameters. We find that common database operations can exercise the full dynamic power range of a server, and that the CPU power consumption of different operators, for the same CPU utilization, can differ by as much as 60%. We also find that for these operations CPU power does not vary linearly with CPU utilization.

Enterprises rely on decision support systems to influence critical business choices. At the same time, IT-related power costs are growing and are a key concern for enterprise executives. Yet, there is little work to date characterizing the power use of decision support systems. Towards this end, we present the first holistic measurements and analysis of an audit-class system running the TPC-H decision support benchmark at the 300GB scale. We first provide a breakdown of the system's power use into its core hardware components. We then explore its power-performance tradeoffs. This investigation shows that there is ample room to improve its energy use without sacrificing much performance. Moreover, the most energy-efficient configuration depends on the workload. These results suggest that, going forward, database software has an important role to play in optimizing for energy use.

Solid state drives perform random reads more than 100x faster than traditional magnetic hard disks, while offering comparable sequential read and write bandwidth. Because of their potential to speed up applications, as well as their reduced power consumption, these new drives are expected to gradually replace hard disks as the primary permanent storage media in large data centers. However, although they may benefit applications that stress random reads immediately, they may not improve database applications, especially those running long data analysis queries. Database query processing engines have been designed around the speed mismatch between random and sequential I/O on hard disks and their algorithms currently emphasize sequential accesses for disk-resident data.

We propose a new paradigm for building scalable distributed systems. Our approach does not require dealing with message-passing protocols, a major complication in existing distributed systems. Instead, developers just design and manipulate data structures within our service called Sinfonia. Sinfonia keeps data for applications on a set of memory nodes, each exporting a linear address space. At the core of Sinfonia is a new minitransaction primitive that enables efficient and consistent access to data, while hiding the complexities that arise from concurrency and failures. Using Sinfonia, we implemented two very different and complex applications in a few months: a cluster file system and a group communication service. Our implementations perform well and scale to hundreds of machines.

As access times to main memory and disks continue to diverge, faster non-volatile storage technologies become more attractive for speeding up data analysis applications. NAND flash is one such promising substitute for disks. Flash offers faster random reads than disk, consumes less power than disk, and is cheaper than DRAM. In this paper, we investigate alternative data layouts and join algorithms suited for systems that use flash drives as the non-volatile store.

Internet applications increasingly rely on scalable data structures that must support high throughput and store huge amounts of data. These data structures can be hard to implement efficiently. Recent proposals have overcome this problem by giving up on generality and implementing specialized interfaces and functionality (e.g., Dynamo [4]). We present the design of a more general and flexible solution: a fault-tolerant and scalable distributed B-tree. In addition to the usual B-tree operations, our B-tree provides some important practical features: transactions for atomically executing several operations in one or more B-trees, online migration of B-tree nodes between servers for load-balancing, and dynamic addition and removal of servers for supporting incremental growth of the system. Our design is conceptually simple. Rather than using complex concurrency and locking protocols, we use distributed transactions to make changes to B-tree nodes. We show how to extend the B-tree and keep additional information so that these transactions execute quickly and efficiently. Our design relies on an underlying distributed data sharing service, Sinfonia [1], which provides fault tolerance and a light-weight distributed atomic primitive. We use this primitive to commit our transactions. We implemented our B-tree and show that it performs comparably to an existing open-source B-tree and that it scales to hundreds of machines. We believe that our approach is general and can be used to implement other distributed data structures easily.

The energy efficiency of computer systems is an important concern in a variety of contexts. In data centers, reducing energy use improves operating cost, scalability, reliability, and other factors. For mobile devices, energy consumption directly affects functionality and usability. We propose and motivate JouleSort, an external sort benchmark, for evaluating the energy efficiency of a wide range of computer systems from clusters to handhelds. We list the criteria, challenges, and pitfalls from our experience in creating a fair energy-efficiency benchmark. Using a commercial sort, we demonstrate a JouleSort system that is over 3.5x as energy-efficient as last year's estimated winner. This system is quite different from those currently used in data centers. It consists of a commodity mobile CPU and 13 laptop drives connected by server-style I/O interfaces.

We propose a new paradigm for building scalable distributed systems. Our approach does not require dealing with message-passing protocols -- a major complication in existing distributed systems. Instead, developers just design and manipulate data structures within our service called Sinfonia. Sinfonia keeps data for applications on a set of memory nodes, each exporting a linear address space. At the core of Sinfonia is a novel minitransaction primitive that enables efficient and consistent access to data, while hiding the complexities that arise from concurrency and failures. Using Sinfonia, we implemented two very different and complex applications in a few months: a cluster file system and a group communication service. Our implementations perform well and scale to hundreds of machines.

Emerging Web services, such as email, photo sharing, and web site archives, must preserve large volumes of quickly accessible data indefinitely into the future. The costs of doing so often determine whether the service is economically viable. We make the case that these applications' demands on large scale storage systems over long time horizons require us to reevaluate traditional system designs. We examine threats to long-lived data from an end-to-end perspective, taking into account not just hardware and software faults but also faults due to humans and organizations. We present a simple model of long-term storage failures that helps us reason about various strategies for addressing some of these threats. Using this model we show that the most important strategies for increasing the reliability of long-term storage are detecting latent faults quickly, automating fault repair to make it cheaper and faster, and increasing the independence of data replicas.

Bugs in complex distributed systems are often hard to find. Many bugs reflect discrepancies between a system's behavior and the programmer's assumptions about that behavior. Differences may be in correctness, in performance characteristics, or both. Our debugging framework, Pip, compares actual behavior with expected behavior and visualizes both. Pip consists of two tools to help reconcile assumptions and actual behavior: an automatic expectations checker and an interactive behavior-explorer GUI.

We present a technique that masks failures in a cluster to provide high availability and fault-tolerance for long-running, parallelized dataflows. We can use these dataflows to implement a variety of continuous query (CQ) applications that require high-throughput, 24x7 operation. Examples include network monitoring, phone call processing, click-stream processing, and online financial analysis. Our main contribution is a scheme that carefully integrates traditional query processing techniques for partitioned parallelism with the process-pairs approach for high availability. This delicate integration allows us to tolerate failures of portions of a parallel dataflow without sacrificing result quality. Upon failure, our technique provides quick fail-over, and automatically recovers the lost pieces on the fly. This piecemeal recovery provides minimal disruption to the ongoing dataflow computation and improved reliability as compared to the straight-forward application of the process-pairs technique on a per dataflow basis. Thus, our technique provides the high availability necessary for critical CQ applications. Our techniques are encapsulated in a reusable dataflow operator called Flux, an extension of the Exchange that is used to compose parallel dataflows. Encapsulating the fault-tolerance logic into Flux minimizes modifications to existing operator code and relieves the burden on the operator writer of repeatedly implementing and verifying this critical logic. We present experiments illustrating these features with an implementation of Flux in the TelegraphCQ code base [8].

The long-running nature of continuous queries poses new scalability challenges for dataflow processing. CQ systems execute pipelined dataflows that may be shared across multiple queries. The scalability of these dataflows is limited by their constituent, stateful operators - e.g. windowed joins or grouping operators. To scale such operators, a natural solution is to partition them across a shared-nothing platform. But in the CQ context, traditional, static techniques for partitioned parallelism can exhibit detrimental imbalances as workload and runtime conditions evolve. Long-running CQ dataflows must continue to function robustly in the face of these imbalances. To address this challenge, we introduce a dataflow operator called flux that encapsulates adaptive state partitioning and dataflow routing. Flux is placed between producer-consumer stages in a dataflow pipeline to repartition stateful operators while the pipeline is still executing. We present the flux architecture, along with repartitioning policies that can be used for CQ operators under shifting processing and memory loads. We show that the flux mechanism and these policies can provide several factors improvement in throughput and orders of magnitude improvement in average latency over the static case.

No abstract available.

We present a continuously adaptive, continuous query (CACQ) implementation based on the eddy query processing framework. We show that our design provides significant performance benefits over existing approaches to evaluating continuous queries, not only because of its adaptivity, but also because of the aggressive cross-query sharing of work and space that it enables. By breaking the abstraction of shared relational algebra expressions, our Telegraph CACQ implementation is able to share physical operators --- both selections and join state --- at a very fine grain. We augment these features with a grouped-filter index to simultaneously evaluate multiple selection predicates. We include measurements of the performance of our core system, along with a comparison to existing continuous query approaches.

No abstract available.

The design and tuning of new access methods (AMs) for nontraditional data types and application areas has always been more of a black art than a rigorous discipline. The designer can only rely on intuition to come up with an effective design; its evaluation and profiling require tedious instrumentation of complex AM code and a host of hand-written scripts. To address these issues, we developed amdb, a visual AM “debugging” tool to support the AM design and implementation process. It is based on the GiST (Generalized Search Tree, [HNP95]) framework for AM construction, which offers the designer an abstracted view of a tree-structured AM and factors out the mechanical aspects of an AM implementation, such as tree traversal, concurrency control and recovery. Amdb is a visual analysis, debugging and profiling tool for AMs that are written as extensions of libgist, a public-domain stand-alone C++ implementation of GiSTs.