A scalable generic transaction model scenario for distributed NoSQL databases

The Journal of Systems and Software 101 (2015) 43–58 Contents lists available at ScienceDirect The Journal of Systems and Software journal homepage: www.elsevier.com/locate/jss A scalable generic transaction model scenario for distributed NoSQL databases Ramesh Dharavath∗, Chiranjeev Kumar Department of Computer Science and Engineering, Indian School of Mines, Dhanbad 826004, Jharkhand, India a r t i c l e i n f o a b s t r a c t Article history: With the development of cloud computing and internet; e-Commerce, e-Business and corporate world rev- Received 21 February 2014 enue are increasing with high rate. These areas not only require scalable and consistent databases but Revised 22 September 2014 also require inter database transaction support. In this paper, we present, a scalable three-tier architecture Accepted 20 November 2014 along with a distributed middle-ware protocol to support atomic transactions across heterogeneous NoSQL Available online 28 November 2014 databases. Our methodology does not compromise on any assumption on the accuracy of failure modalities. Keywords: Hence, it is suitable for a class of heterogeneous distributed systems. To achieve such a target, our architec- Atomic transaction tural model exploits an innovative methodology to achieve distributed atomic transactions. We simulate this Column-oriented distributed databases architectural setup with different latency tests under different environments to produce reliable impact and NoSQL databases correctness. © 2014 Elsevier Inc. All rights reserved. 1. Introduction arguments that call for the semi structured data (Abiteboul et al., 1999). At the same time, the full power of relational models with mul- Distributed computing infrastructures were emerged as the de- tifaceted transactions and queries would not require in some perspec- ployment platforms for a variety of applications. Increasingly, desk- tives, where simple operations are quite enough (Stonebraker and top and mobile applications are using distributed infrastructure to Cattel, 2011). Also, in some instances, ACID consistency and complete take advantage of high scalability characteristics. Massive usage of consistency assured by relational model is not essential. It has been internet produces huge amount of data. These vast volumes of data observed that many internet application areas related to networking are also called Big Data. This information can be used in many ways. domain need both scalability and flexibility in construction, while This Big Data is beyond the capability of traditional databases. New being fulfilled with operations and weak forms of consistency. With type of databases have been introduced to deal with this Big Data, these drives, a number of new systems are not following the RDBMS these are called column oriented databases and also known as NoSQL models. Their futures are limited and support simple operations with which are scalable, eventual consistent, and reliable. Due to these limited flexibility in the structure of data. According to McKinsey features NoSQL databases like Big-Table (Chang et al., 2006, 2008), Global Institute (Manyika et al., 2011), the data volume is growing at Cassandra (Lakshman and Malik, 2010), HBase (Vora, 2011), Mon- a rate of 40% per year. From this, we conclude that NoSQL is better for goDB (Membrey et al., 2010), and Amazon S3 (DeCandia et al., 2007) large data volume applications, where the date rate increases rapidly. are successfully used by almost all internet giants. NoSQL databases There is wide variety of systems with NoSQL (Cattel, 2010; Chang offer great scalability and availability for applications. et al., 2008) exposes a different interface models and APIs. Certainly, Relational database systems (RDBMSs) also dominate the environ- it has been pointed that, the lack of standard is a great trepidation for ment by providing set of services which refer variety of requirements. organizations in adopting any of these systems (Stonebraker, 2011), But, these require analytical processing and decision support. How- applications, and data are not portable and reusable with others. Also, ever, some trepidation has recently materialized towards RDBMSs. each of these systems has specific objectives, and it is customized on First, it has been argued that, there are cases where RDBMSs per- few specific prototypes. As an outcome, it might be possible that com- formances are not passable. Second, the structure of the model is plicated applications could benefit from the use of NoSQL systems. considered to be too rigid and not useful in some instances with the Traditional databases are useful for those applications, where consis- tency and transaction support is needed in more suitable fashion. But, the current trend applications such as corporate world, e-business ∗ Corresponding author. Tel.: +91 326 2235795; fax: +91 326 2296563/2296613. and e-commerce needs database support which is not only scalable E-mail addresses: [email protected] (D. Dharavath), [email protected] as well as consistent too. With the massive success of NoSQL databases (C. Kumar). and growth of e-commerce, e-banking and corporate data, we assume http://dx.doi.org/10.1016/j.jss.2014.11.037 0164-1212/© 2014 Elsevier Inc. All rights reserved. 44 D. Dharavath, C. Kumar / The Journal of Systems and Software 101 (2015) 43–58 that, in future, NoSQL databases can be used in e-commerce and e- on this, many approaches have been suggested to avoid serializa- banking areas. In order to preserve atomicity and isolation related in- tion anomalies. These approaches include static analysis (Jorwekar stances, in a previous work (Ramesh et al., 2012a,b, 2013), we have im- et al., 2007) as well as runtime detection of anomalies (Bornea plemented an architectural methodology for multi row transactions et al., 2011; Cahill et al., 2009; Revilak et al., 2011). Methodology on column oriented distributed databases using RDBMS. We have also presented by (Bornea et al., 2011; Cahill et al., 2009; Jung et al. 2011) proposed a suitable methodology that preserves accuracy of atomic tend to be distrustful and can lead to unnecessary aborts. The PSSI transactions in heterogeneous distributed column oriented database approach (Revilak et al., 2011) avoids such problems and aborts only environment (Dharavath et al., 2014). The observations above have the transactions that lead to serialization anomalies. However, these motivated us to look for tools and methods that can improve the con- approaches were developed in the context of traditional relational sequences of the heterogeneity in distributed NoSQL database sys- databases and (except in Jung et al., 2011) provided solutions only for tems and can also enable interoperability between them. centralized databases. Thomson et al. (2012) proposed a Calvin layer As a first step in this direction, in this paper, we present here a “dis- to support fast and scalable transactions. Calvin layer is concerned tributed three-tier architecture” as well as “distributed middleware pro- about replication and scheduling. Calvin has no single point failure. tocol” to support atomic transaction across heterogeneous distributed This transaction model is useful to get high throughput but replica- NoSQL databases. Proposed framework is entirely distributed; hence tion relaxed ACID is an issue with this layer. Another method called it is infinitely scalable, highly available, very flexible and cost effective. MAV (Monotonic Atomic View) proposed by Peter et al. (2013), which We also exemplify the correctness and effectiveness of our framework requires disallowing reading intermediate writes of a transaction. But in different aspects. it incurs two writes for every client side write and achieves 75% of the throughput only. Han et al. (2011), made a survey on NoSQL database 2. Related literature that explains distinct column-oriented databases like MongoDB, Cas- sandra, Hyper table etc. This paper also describes about data model Frolund and Guerraoui (2001, 2002), have recently proposed a reli- used by NoSQL and explores features of NoSQL. In this paper, we de- ability framework called e-transaction (exactly once transaction), that scribe a framework to provide atomicity for distributed transactions. provide atomicity for distributed transaction. An e- transaction frame- Proposed framework is scalable and flexible in nature and recovers work is based on three tier architecture and protocol implementation from failure instances. In novelty, this framework fulfils present as for each tier is clearly explained. In this framework, there is a chance well as future requirements related to real world applications. of single point failure and scalability issue. There have been several works that actively replicates database systems in this way. Pacitti 3. Proposed methodology et al. (2003), Whitney et al. (1997), Stonebraker et al. (2007), and Jones et al. (2010) proposed about performance of transaction pro- In this work, we describe and present a distributed architecture cessing in a distributed database without concurrency control mech- (depicted in Fig. 1) to provide atomic transaction instance across het- anism. Moreover, these methodologies execute the transactions se- erogeneous distributed column-oriented NoSQL databases. This ar- rially which is equivalent to a serial order i.e. where a node can be chitecture is scalable and flexible in transactional issues. This archi- a single CPU core in a multi-core server (Stonebraker et al., 2007). tecture is mainly inspired by Apache HBase data-store architecture By executing the transactions serially, no determinism due to thread (Lars, 2011). In HBase (Vora, 2011), master nodes are managed by scheduling of concurrent transactions is eliminated, and active repli- zookeeper. Master is used for load balancing and managing region cation is easier to achieve. However, serializing transactions can limit nodes. In the same way in our architecture, manager nodes are man- transactional throughput, since if a transaction stalls (e.g. for a net- aged by zookeeper and manager is used for balancing the load among work read), other transactions are unable to take over. labour nodes. HBase is having the control to store and retrieve Big- Other researchers have proposed techniques for providing multi- Data, in other hand our architecture is used to provide atomicity for row transactions (Peng and Dabek, 2010), (Zhang and Sterck, 2010). transaction across heterogeneous distributed NoSQL databases. As The Percolator system (Peng and Dabek, 2010) addresses the problem name implies, three-tier architecture is divided physical entities in of providing SI-based transaction for Big- Table (Chang et al., 2008). three layers: (i) client (ii) middle-tier, and (iii) distributed database However, it does not address the problem of ensuring serializabil- (HBase). ity of transactions. Moreover, it is intended for offline processing of data. The work presented in (Zhang and Sterck, 2010) does not ade- 3.1. Client quately address issues related to recovery and robustness when some transaction fails. Kallman et al. (2008) proposed a technique which is Architecture supports multiple client nodes denoted by C1 , C2 . . . related to main memory transaction processing. This technique has Cp . User interacts with the entire architecture by utilizing the not explored in scaling the performance of transactions which were situated in main memory. Dadiomov et al. (2003), patents a system for Mn M1 M2 distributed transaction processing with asynchronous message deliv- … Active ery. According to this, a transaction is combination of many database operations. These operations spans over many distributed databases. All operations in a transaction must be performed atomically. To im- Zookeeper plement this system authors use message queue (MQ) concept. This Client MySQL message queue (MQ) guarantee exactly-once in-order message de- Cluster livery. For each transaction there is a coordinator. With the help of message queue (MQ) message is passed between coordinator and Heterogeneous participating databases. In this proposed system, atomicity can be NoSQL achieved using two-phase commit (2PC) protocol. databases The problems of transaction serializability have been extensively L1 L2 … Lm studied in the form of different models (Fekete et al., 2005; Bornea et al., 2011; Cahill et al., 2009; Revilak et al., 2011; Jung et al., 2011). Middle-Tier The work presented by Fekete et al. (2005) exemplifies about nec- essary conditions for non-serializable transaction executions. Based Fig. 1. Schematized three-tier architecture.