MonetDB/DataCell: Online Analytics in a Streaming Column-Store Erietta Liarou, Stratos Idreos, Stefan Manegold, Martin Kersten CWI, Amsterdam, The Netherlands [email protected] [email protected] [email protected] [email protected] ABSTRACT tion, web log analysis requires fast analysis of big streaming In DataCell, we design streaming functionalities in a mod- data for decision support. ern relational database kernel which targets big data analyt- A new processing paradigm is born [15, 19, 11] where in- ics. This includes exploitation of both its storage/execution coming streams of data need to quickly be analyzed and engine and its optimizer infrastructure. We investigate the possibly combined with existing data to discover trends and opportunities and challenges that arise with such a direction patterns. Subsequently, the new data may also enter the and we show that it carries significant advantages for mod- data warehouse and be stored as normal for further analy- ern applications in need for online analytics such as web sis if necessary. This new paradigm requires scalable query logs, network monitoring and scientific data management. processing that can combine continuous querying for fast The major challenge then becomes the efficient support for reaction to incoming data with traditional querying for ac- specialized stream features, e.g., multi-query processing and cess to the existing data. However, neither pure database incremental window-based processing as well as exploiting technology nor pure stream technology are designed for this standard DBMS functionalities in a streaming environment purpose. Database systems do not qualify for continuous such as indexing. query processing, while data stream systems are not built In this demo, we present the DataCell system, an exten- to scale for big data analysis. sion of the MonetDB open-source column-store for online an- DataCell Motivation. MonetDB/DataCell aims to pro- alytics. The demo gives the user the opportunity to experi- vide scalable online analytics. We begin from a state of ence the features of DataCell such as processing both stream the art column-store design for big data analytics, Mon- and persistent data and performing window based process- etDB, and extend it with online functionality. The goal is ing. The demo provides a visual interface to monitor the to fully exploit the generic storage and execution engine of critical system components, e.g., how query plans transform the DBMS as well as its complete optimizer stack. Stream from typical DBMS query plans to online query plans, how processing then becomes primarily a query scheduling task, data flows through the query plans as the streams evolve, i.e., make the proper queries see the proper portion of stream how DataCell maintains intermediate results in columnar data at the proper time. A positive side-effect is that our form to avoid repeated evaluation of the same stream por- architecture supports SQL'03, which allows stream applica- tions, etc. The demo also provides the ability to interac- tions to exploit sophisticated query semantics. Numerous tively set the test scenarios regarding input data and various research and technical questions immediately arise. The DataCell knobs. most prominent issues are the ability to provide specialized stream functionality and hindrances to guarantee real-time constraints for event handling. 1. INTRODUCTION Contributions and Demo. Paper [15] illustrates the Numerous applications nowadays require online analytics basic DataCell architecture and sets the research path and over high rate streaming data. For example, emerging ap- critical milestones. In addition, DataCell ships together plications over mobile data can exploit the big mobile data with the MonetDB open-source system. Here, we present streams for advertising and traffic control. In addition, the a demo based on MonetDB/DataCell. The demo showcases recent and continuously expanding massive cloud infrastruc- several of the key aspects of the MonetDB/DataCell design tures require continuous monitoring to remain in good state such as the ability to do both stream processing and normal and prevent fraud attacks. Similarly, scientific databases queries and the ability to provide window based processing. create data at massive rates daily or even hourly. In addi- Through a graphical user interface the user of the demo can observe the query execution inside DataCell, e.g., how the various columnar structures are populated, how intermedi- Permission to make digital or hard copies of all or part of this work for ate results are kept around to avoid reevaluation for sliding personal or classroom use is granted without fee provided that copies are window queries and how the shape of a normal query plan not made or distributed for profit or commercial advantage and that copies changes to a continuous query plan through the optimizer. bear this notice and the full citation on the first page. To copy otherwise, to In addition, the user can interact with the system, posing republish, to post on servers or to redistribute to lists, requires prior specific continuous or one-time queries or choosing some pre-defined permission and/or a fee. Articles from this volume were invited to present scenarios as well as varying both input parameters and Da- their results at The xth International Conference on Very Large Data Bases. Proceedings of the VLDB Endowment, Vol. X, No. Y taCell knobs. Copyright 20xy VLDB Endowment 2150-8097/11/XX... $ 10.00. Paper Organization. The rest of the paper is organized Receptors Parser/Compiler Emitters as follows. Section 2 briefly discusses related work while Sec- tion 3 gives an overview of the main design s of DataCell. Optimizer Then, Section 4 presents the demonstration scenarios as well Rewriter as the ways the user can interact with the system and how the demo visualizes query execution and system status. Fi- Scheduler/Factories nally, Section 5 concludes the paper. 2. RELATED WORK DataCell fundamentally differs from existing stream ef- Kernel forts [2, 3, 4, 7, 8, 9, 12, 21, 5, 10], etc. by building on top Baskets Tables of the storage and execution engine of a DBMS kernel. It opens a very interesting path towards exploiting and merg- ing technologies from both worlds. Figure 1: The DataCell Architecture Compared to even earlier efforts on active databases, e.g., [20], DataCell adds support for specialized stream function- alities, such as incremental processing. Incremental process- each attribute. Advanced column-stores process one column ing in a DBMS has been studied in the context of updat- at a time, using late tuple reconstruction, discussed in, e.g., ing materialized views, e.g., [6, 13], but there the scenario is [1, 14]. Intermediates are also in column format. This allows very different given that it targets read-mostly environments the engine to exploit CPU- and cache-optimized vector-like whereas an online scenario is by definition a write-only one. operator implementations throughout the whole query eval- Truviso Continuous Analytics system [11], a commercial uation, using an efficient bulk processing model instead of product of Truviso, is another recent example that follows the typical tuple-at-a-time volcano approach. This way, a the same approach as DataCell. They extend the open select operator for example, operates on a single column, fil- source PostgreSQL database [18] to enable continuous anal- tering the qualifying values and producing an intermediate ysis of streaming data, tackling the problem of low latency that holds their tuple IDs. This intermediate can then be query evaluation over massive data volumes. TruCQ inte- used to retrieve the necessary values from a different column grates streaming and traditional relational query process- for further actions, e.g., aggregations, further filtering, etc. ing in such a way that ends-up to a stream-relational da- The key point is that in DataCell these intermediates can be tabase architecture. It is able to run SQL queries continu- exploited for flexible incremental processing strategies, i.e., ously and incrementally over data while they are still com- we can selectively keep around the proper intermediates at ing and before they are stored in active database tables (if the proper places of a plan for efficient future reuse. they). TruCQ's query processing significantly outperforms DataCell. DataCell [15] is positioned between the SQL traditional store-first-query-later database technologies as compiler/optimizer and the DBMS kernel. The SQL com- the query evaluation has already been initiated when the piler is extended with a few orthogonal language constructs first tuples arrive. It allows evaluation of one-time queries, to recognize and process continuous queries. The query plan continuous queries, as well as combinations of both types. as generated by the SQL optimizer is rewritten to a contin- Another recent work, coming from the HP Labs [19], also uous query plan and handed over to the DataCell scheduler. confirms the strong research attraction for this trend. It In turn, the scheduler handles the execution of the plan. defines an extended SQL query model that unifies queries Figure 1 shows a DataCell instance. It contains receptors over both static relations and dynamic streaming data, by and emitters, i.e., a set of separate processes per stream and developing techniques to generalize the query engine. It also per client, respectively, to listen for new data and to deliver extends the PostgreSQL database kernel [18], building an results. They form the edges of the architecture and the engine that can process persistent and streaming data in bridges to the outside world, e.g., to sensor drivers. a uniform design. First, they convert the stream into a Baskets/Columns. The key idea is that when an event sequence of chunks and then continuously call the query stream enters the system via a receptor, stream tuples are over each sequential chunk. The query instance never shuts immediately stored in a lightweight table, called basket.