An instant-search engine computes answers immediately as a user types a query character by character. In this paper, we study how to systematically collect information about user behaviors when they interact with an instant search engine, especially in a real-time environment. We present a solution, called RILCA, which uses front-end techniques to keep track of rich information about user activities. This information provides more insights than methods based on traditional Web servers such as Apache. We store the log records in a relational DBMS system, and leverage the existing powerful capabilities of the DBMS system to analyze the log records efficiently. We study how to use a dashboard to monitor and analyze log records in real time. We conducted experiments on real data sets collected from two live systems to show the benefits and efficiency of these techniques.
Automating predictive machine learning entails the capability of properly triggering the update of the trained models. To this aim, the degradation of predictive models has to be continuously evaluated over time to detect data distribution drifts between the original training set and the new data. Traditionally, prediction performance is used as a degradation metric. However, prediction quality indices require ground-truth class labels to be known for the newly classified data, making them unsuitable for real-time applications, as ground-truth labels might be totally absent or be available only later. In this paper, we propose a novel unsupervised methodology to automatically detect prediction-quality degradation of machine learning models. Thanks to the unsupervised approach and a novel scalable estimation technique, we provide an effective and efficient solution to the above-mentioned problem with soft real-time constraints. Specifically, our approach is able to detect class-based concept drift, i.e., when new data contain samples that do not fit the set of class labels known by the currently-trained predictive model. Experiments on synthetic and real-world public datasets show the effectiveness of the proposed methodology in automatically detecting and describing concept drift caused by changes in the class-label data distributions. Thanks to its scalability performance, the proposed approach is suitable for soft real-time applications such as predictive maintenance, Industry 4.0, and text mining.
Contemporary end-to-end data pipelines need to combine many diverse workloads such as machine learning, relational operations, stream dataflows, tensor transformations, and graphs. For each of these workload types, there exists several frontends (e.g., SQL, Beam, Keras) based on different programming languages as well as different runtimes (e.g., Spark, Flink, Tensorflow) that optimize for a particular frontend and possibly a hardware architecture (e.g., GPUs). The resulting pipelines suffer in terms of complexity and performance due to excessive type conversions, materialization of intermediate results, and lack of cross-framework optimizations. Arcon aims to provide a unified approach to declare and execute tasks across frontend-boundaries as well as enabling their seamless integration with event-driven services at scale. In this demonstration, we present Arcon and through a series of use-case scenarios demonstrate that its execution model is powerful enough to cover existing as well as upcoming real-time computations for analytics and application-specific needs.
We explore the performance and resource trade-offs of two alternative approaches to streaming state management. When the state size exceeds the amount of available memory, systems can either scale out and partition the state across distributed computing nodes or rely on secondary storage and divide the state into 'hot' and 'cold' sets. Scaling out a streaming computation might introduce coordination overhead among parallel workers, while flushing state to disk requires efficient data structures and careful caching policies to minimise expensive I/O. To study the characteristics of these state management approaches, we present an integration of the Timely Dataflow stream processing engine with the FASTER embedded key-value store. We demonstrate a prototype that allows users to transparently maintain arbitrary larger-than-memory state with low overhead by making only minimal changes to application code. Our preliminary experimental results show that managed state incurs acceptable overhead over built-in in-memory data structures and, in some cases, performs better when relying on secondary storage in a single node as opposed to scaling out to multiple nodes.
Traditional row store has the disadvantage of reading irrelevant attributes into main memory, when only a few attributes are queried and it incurs a large number of cache misses. Column store, while being cache efficient, suffers from large "stitching" costs if the number of attributes queried are large. In this paper, we introduce FAST, an intuitive, hybrid combination of both these storage mechanisms that leverages the main memory to reduce the I/O time and cache misses to efficiently serve ad-hoc read-only queries in real time. Our system acts as a commercial off-the-shelf (COTS) solution on top of existing databases. We show that on average, FAST executes queries up-to an order of magnitude faster than row storage and as much as twice as fast than column storage for an ad-hoc data warehouse workload. Our results show superior performance of FAST on multiple TPC-H queries. We also present techniques to automatically adapt the main memory layout to a changing workload.
Large-scale classification of text streams is an essential problem that is hard to solve. Batch processing systems are scalable and proved their effectiveness for machine learning but do not provide low latency. On the other hand, state-of-the-art distributed stream processing systems are able to achieve low latency but do not support the same level of fault tolerance and determinism. In this work, we discuss how the distributed streaming computational model and fault tolerance mechanisms can affect the correctness of text classification data flow. We also propose solutions that can mitigate the revealed pitfalls.
Many modern applications require realtime processing of large volumes of high speed data. Such data processing can be specified in the form of a dataflow graph that exposes multiple opportunities for parallelizing its execution, in the form of data, pipeline and task parallelism. This paper focuses on the problem of effectively parallelizing ordered streaming computations by exploiting low overhead, shared memory parallelism on a multicore machine. We propose an adaptive runtime that dynamically maps the exposed parallelism in a streaming computation to that of the machine using scheduling heuristics. Further, we address some key problems in effectively realizing ordered data parallelism. We propose a new approach for parallelizing partitioned stateful operators that can handle load imbalance across partitions effectively and mostly avoid delays due to ordering constraints. We also present a low latency, non-blocking concurrent data structure to order outputs produced by concurrent workers on an operator. Finally, we perform an in-depth empirical evaluation illustrating the trade-offs and effectiveness of our concurrent data-structures and scheduling heuristics using micro-benchmarks and on the TPCx-BB benchmark.
The need for real-time analysis is still spreading and the number of available streaming sources is increasing. The recent literature has plenty of works on Data Stream Processing (DSP). In a streaming environment, the data incoming rate varies over time. The challenge is how to efficiently deploy these applications in a cluster. Several works have been conducted on improving the latency of the system or to minimize the allocated resources per application through time. However, to the best of our knowledge, none of the existing works takes into consideration the user needs for a specific application, which is different from one user to another. In this paper, we propose Moria, a goal-oriented framework for dynamically optimizing the resource allocation built on top of Apache Flink. The system takes actions based on the user application and on the incoming data characteristics (i.e., input rate and window size). Starting from an initial estimation of the resources needed for the user query, at each iteration we improve our cost function with the collected metrics from the monitored system about the incoming data, to fulfill the user needs. We present a series of experiments that show in which cases our dynamic estimation outperforms the baseline Apache Flink and the thumb rule estimation alone performed at the deployment of the applications.
Complex data analysis scenarios often require discovering and combining multiple data sources. Data scientists usually formulate a series of SQL queries building on each other, also called a session, to iteratively derive results. However, due to a lack of familiarity with data sources or the complexity of query results, it can be a hard task to decide on the next query iteration solely based on the results of the last one. While existing approaches provide mechanisms to assess the results of a specific query, support for analyzing results in the context of the respective session remains mostly absent. Such approaches do also not seamlessly integrate with established tools and workflows. To overcome these problems, we introduce OCEANProfile, a framework for session-based profiling of query results. Query results are intercepted at driver level and streamed into our framework for automated data profiling. Result profiles can be compared with those of previous queries and visualized in a companion app compatible with existing analysis tools. Visualizations are automatically ranked according to their usefulness in the context of the respective session.
As enterprise information systems are collecting event streams from various sources, the ability of a system to automatically detect anomalous events and further provide human readable explanations is of paramount importance. In this position paper, we argue for the need of a new type of data stream analytics that can address anomaly detection and explanation discovery in a single, integrated system, which not only offers increased business intelligence, but also opens up opportunities for improved solutions. In particular, we propose a two-pass approach to building such a system, highlight the challenges, and offer initial directions for solutions.
Software organizations want to be able to base their decisions on the latest set of available data and the real-time analytics derived from them. In order to support "real-time enterprise" for software organizations and provide information transparency for diverse stakeholders, we integrate heterogeneous data sources about software analytics, such as static code analysis, testing results, issue tracking systems, network monitoring systems, etc. To deal with the heterogeneity of the underlying data sources, we follow an ontology-based data integration approach in this paper and define an ontology that captures the semantics of relevant data for software analytics. Furthermore, we focus on the integration of such data sources by proposing two approaches: a static and a dynamic one. We first discuss the current static approach with a predefined set of analytic views representing software quality factors and further envision how this process could be automated in order to dynamically build custom user analysis using a semi-automatic platform for managing the lifecycle of analytics infrastructures.
Stream processing has emerged as a paradigm for applications that require low-latency evaluation of operators over unbounded sequences of data. Defining the semantics of stream processing is challenging in the presence of distributed data sources. The physical and logical order of data in a stream may become inconsistent in such a setting. Existing models either neglect these inconsistencies or handle them by means of data buffering and reordering techniques, thereby compromising processing latency. In this paper, we introduce the Dual Streaming Model to reason about physical and logical order in data stream processing. This model presents the result of an operator as a stream of successive updates, which induces a duality of results and streams. As such, it provides a natural way to cope with inconsistencies between the physical and logical order of streaming data in a continuous manner, without explicit buffering and reordering. We further discuss the trade-offs and challenges faced when implementing this model in terms of correctness, latency, and processing cost. A case study based on Apache Kafka illustrates the effectiveness of our model in the light of real-world requirements.
Oracle TimesTen Scaleout is a shared-nothing scale-out inmemory database designed for extreme OLTP workloads, such as IoT, real-time fraud detection, telecommunications etc. TimesTen Scaleout features rich SQL including complex queries with joins, aggregations and analytic functions, transparent distributed execution, full ACID multi-statement transactions, global secondary indexes and sequences. The design features built-in high availability using a k-safe data duplication mechanism, and transparent failure handling in order to minimize application down time. All management functions such as installation, configuration, and monitoring are provided via a centralized management repository. We describe some of the challenges in developing such a scale-out in-memory database architecture and show extreme performance results exceeding 100 million transactions per second and 1 billion SQL selects per second.
In the new digital economy, on demand access of real time enterprise data is critical to modernize cross organizational, cross partner, and online consumer functions. In addition to on premise legacy data, enterprises are producing an enormous amount of real-time data through new hybrid cloud applications; these event streams need to be collected, transformed and analyzed in real-time to make critical business decision. Traditional Extract-Load-Transform (ETL) processes are no longer sufficient and need to be re-architected to account for streaming, heterogeneity, usability, extensibility (custom processing), and continuous validity. Striim is a novel end-to-end distributed streaming ETL and intelligence platform that enables rapid development and deployment of streaming applications. Striim's real-time ETL engine has been architected from ground-up to enable both business users and developers to build and deploy streaming applications. In this paper, we describe some of the core features of Striim's ETL engine (i) built-in adapters to extract and load data in real-time from legacy and new cloud sources/targets (ii) an extensible SQL-based transformation engine to transform events; users can inject custom logic via a component called Open Processor (iv) New primitives like MODIFY, BEFORE and AFTER and (v) built-in data validation that continuously checks if everything is continually making it to the destination.
Pure machine-based solutions usually struggle in the challenging classification tasks such as entity resolution (ER). To alleviate this problem, a recent trend is to involve the human in the resolution process, most notably the crowdsourcing approach. However, it remains very challenging to effectively improve machine-based entity resolution with limited human effort. In this paper, we investigate the problem of human and machine cooperation for ER from a risk perspective. We propose to select the machine-labeled instances at high risk of being mislabeled for manual verification. For this task, we present a risk model that takes into consideration the human-labeled instances as well as the output of machine resolution. Finally, we evaluate the performance of the proposed risk model on real data. Our experiments demonstrate that it can pick up the mislabeled instances with considerably higher accuracy than the existing alternatives. Provided with the same amount of human cost budget, it can also achieve better resolution quality than the state-of-the-art approach based on active learning.
In current Big Data scenarios, traditional data warehousing and Online Analytical Processing (OLAP) operations on cubes are clearly not sufficient to address the current data analysis requirements. Nevertheless, OLAP operations and models can expand the possibilities of graph analysis beyond the traditional graph-based computation. In spite of this, there is not much work on the problem of taking OLAP analysis to the graph data model. In previous work we proposed a multidimensional (MD) data model for graph analysis, that considers not only the basic graph data, but background information in the form of dimension hierarchies as well. The graphs in our model are node-and edge-labelled directed multi-hypergraphs, called graphoids, defined at several different levels of granularity. In this paper we show how we implemented this proposal over the widely used Neo4J graph database, discuss implementation issues, and present a detailed case study to show how OLAP operations can be used on graphs.
More and more organizations (commercial, health, government and security) currently base their decisions on real-time analysis of fast arriving, large volumes of data streams. For such analysis to lead to actionable information in real-time and at the right time, the most recent data needs to be processed within a specified delay target. Effective solutions for analysis of such data streams rely on two techniques, (1) incremental sliding-window computation of aggregates, to avoid unnecessary recomputations and (2) intelligent scheduling of computational steps and operations. In this paper, we propose a solution that combines both of these techniques to find highly correlated data streams in real-time, using the Pearson Correlation Coefficient as a correlation metric for two windows of data streams. Specifically, we propose to partition a set of data streams into micro-batches that capture the delay target, use sliding windows within a range as the subsequences of values exhibiting a certain level of correlation, utilize the idea of sufficient statistics to incrementally compute the Pearson Correlation Coefficient of pairs of sliding windows, and adopt a deadline-aware priority scheduling to detect the highly correlated pairs of data streams. Our experimental results show that our scheme and in particular our Price-DCS with warm start scheduling algorithm outperform existing ones and enable high degree of interactivity in correlating live data streams micro-batches.
Integrating low-latency data streaming into data warehouse architectures has become an important enhancement to support modern data warehousing applications. In these architectures, heterogeneous workloads with data ingestion and analytical queries must be executed with strict performance guarantees. Furthermore, the data warehouse may consists of multiple different types of storage engines (a.k.a., polystores or multi-stores). A paramount problem is data placement; different workload scenarios call for different data placement designs. Moreover, workload conditions change frequently. In this paper, we provide evidence that a dynamic, workload-driven approach is needed for data placement in polystores with low-latency data ingestion support. We study the problem based on the characteristics of the TPC-DI benchmark in the context of an abbreviated polystore that consists of S-Store and Postgres.
A telecommunication company (telco) is traditionally only perceived as the entity that provides telecommunication services, such as telephony and data communication access to users. However, the IP backbone infrastructure of such entities spanning densely urban spaces and widely rural areas, provides nowadays a unique opportunity to collect immense amounts of mobility data that can provide valuable insights for road traffic management and avoidance. In this paper we outline the components of the Traffic-TBD (Traffic Telco Big Data) architecture, which aims to become an innovative road traffic analytic and prediction system with the following desiderata: i) provide micro-level traffic modeling and prediction that goes beyond the current state provided by Internet-based navigation enterprises utilizing crowdsourcing; ii) retain the location privacy boundaries of users inside their mobile network operator, to avoid the risks of exposing location data to third-party mobile applications; and iii) be available with minimal costs and using existing infrastructure (i.e., cell towers and TBD data streams are readily available inside a telco). Road traffic understanding, management and analytics can minimize the number of road accidents, optimize fuel and energy consumption, avoid unexpected delays, contribute to a macroscopic spatio-temporal understanding of traffic in cities but also to "smart" societies through applications in city planning, public transportation, logistics and fleet management for enterprises, startups and governmental bodies.
