An epidemic approach to dependable key-value substrates

2014 IEEE 33rd International Symposium on Reliable Distributed Systems

Vilaça

et al. 2014

Self Cite

Very large scale distributed systems provide some of the most interesting research challenges while at the same time being increasingly required by nowadays applications. The escalation in the amount of connected devices and data being produced and exchanged, demands new data management systems. Although new data stores are continuously being proposed, they are not suitable for very large scale environments. The high levels of churn and constant dynamics found in very large scale systems demand robust, proactive and unstructured approaches to data management. In this paper we propose a novel data store solely based on epidemic (or gossip-based) protocols. It leverages the capacity of these protocols to provide data persistence guarantees even in highly dynamic, massive scale systems. We provide an open source prototype of the data store and correspondent evaluation.

Section: System Designmentioning

confidence: 99%

DATAFLASKS: Epidemic Store for Massive Scale Systems

2014 IEEE 33rd International Symposium on Reliable Distributed Systems

Vilaça

et al. 2014

Self Cite

“…This upper layer provides 1) client interface, 2) caching, 3) concurrency control, and 4) high level processing. As described in [7], even though DATADROPLETSis itself decentralized, we assume this layer to run mainly in memory and in a moderate size environment. This allows DATADROPLETS to be based on a structured approach.…”

Section: Stratusmentioning

confidence: 99%

“…We posit that an unstructured but resilient approach to data management is more appropriate in the context of such largescale systems. In [7], we proposed a novel two-layer approach to the design of a key-value data store. This two-layer approach separates client interface and concurrency control (top layer) from the actual data storage (bottom layer).…”

Section: Introductionmentioning

confidence: 99%

DATAFLASKS: An epidemic dependable key-value substrate

2013 43rd Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN)

Vilaça

et al. 2013

Self Cite

Recently, tuple-stores have become pivotal structures in many information systems. Their ability to handle large datasets makes them important in an era with unprecedented amounts of data being produced and exchanged. However, these tuple-stores typically rely on structured peer-to-peer protocols which assume moderately stable environments. Such assumption does not always hold for very large scale systems sized in the scale of thousands of machines. In this paper we present a novel approach to the design of a tuple-store. Our approach follows a stratified design based on an unstructured substrate. We focus on this substrate and how the use of epidemic protocols allow reaching high dependability and scalability.

“…Partitioning the set of nodes into k several groups of increasing uptime, allows to assign critical services to more stable nodes, and less critical services to less stable ones. Examples include assigning privileged roles to more stable nodes to improve the quality of a streaming application [18], or allocating a data partition to a group of nodes in a key-value store [11]. The operation of partitioning in k groups according to node-specific criteria is called distributed slicing [6,9,13].…”

Section: Introductionmentioning

confidence: 99%

“…Undesired slice changes or oscillations between two slices tend to appear more frequently for nodes that lie at the "borders" of slices, that is, at the boundary of slices in the virtual ranking of all attributes. For instance, in the key-value store application mentioned above [11], a slice change results in discarding a potentially large fraction of hard state for the current slice and getting the new state from nodes of the new slice, which can be costly.…”

Section: Introductionmentioning

confidence: 99%

Slead: Low-Memory, Steady Distributed Systems Slicing

Distributed Applications and Interoperable Systems

Rivière

et al. 2012

Self Cite

Abstract. Slicing a large-scale distributed system is the process of autonomously partitioning its nodes into k groups, named slices. Slicing is associated to an order on node-specific criteria, such as available storage, uptime, or bandwidth. Each slice corresponds to the nodes between two quantiles in a virtual ranking according to the criteria.For instance, a system can be split in three groups, one with nodes with the lowest uptimes, one with nodes with the highest uptimes, and one in the middle. Such a partitioning can be used by applications to assign different tasks to different groups of nodes, e.g., assigning critical tasks to the more powerful or stable nodes and less critical tasks to other slices.Assigning a slice to each node in a large-scale distributed system, where no global knowledge of nodes' criteria exists, is not trivial. Recently, much research effort was dedicated to guaranteeing a fast and correct convergence in comparison to a global sort of the nodes.Unfortunately, state-of-the-art slicing protocols exhibit flaws that preclude their application in real scenarios, in particular with respect to cost and stability. In this paper, we identify steadiness issues where nodes in a slice border constantly exchange slice and large memory requirements for adequate convergence, and provide practical solutions for the two. Our solutions are generic and can be applied to two different state-of-the-art slicing protocols with little effort and while preserving the desirable properties of each. The effectiveness of the proposed solutions is extensively studied in several simulated experiments.