Consistency-based service level agreements for cloud storage

Abstract

Choosing a cloud storage system and specific operations for reading and writing data requires developers to make decisions that trade off consistency for availability and performance. Applications may be locked into a choice that is not ideal for all clients and changing conditions. Pileus is a replicated key-value store that allows applications to declare their consistency and latency priorities via consistencybased service level agreements (SLAs). It dynamically selects which servers to access in order to deliver the best service given the current configuration and system conditions. In application-specific SLAs, developers can request both strong and eventual consistency as well as intermediate guarantees such as read-mywrites. Evaluations running on a worldwide test bed with geo-replicated data show that the system adapts to varying client-server latencies to provide service that matches or exceeds the best static consistency choice and server selection scheme. IntroductionCloud storage systems, such as the currently popular class of "NoSQL" data stores, have been designed to meet the needs of diverse applications from social networking to electronic commerce. Such storage services invariably replicate application data on multiple machines to make it highly available. Many provide a relaxed form of consistency, eventual consistency, in order to achieve elastic scalability and good performance while some strive for strong consistency to maintain the semantics of one-copy serializability. To allow local access and ensure data survivability even during a complete datacenter failure, many storage systems offer "geo-replication," the option of replicating data across different regions of the world.With data being replicated on a worldwide scale, the inherent trade-offs between performance and consistency are accentuated due to the high communication latencies between datacenters. The performance difference between reads with different consistencies can be substantial. This is not surprising. Strongly consistent reads generally involve multiple replicas or must be serviced by a primary replica, whereas eventually consistent reads can be answered by the closest replica. Even within a datacenter, the latency of strongly consistent reads has been measured as eight times that of reads with weaker session guarantees [26]. With geo-replication, our studies show that the performance can differ by more than two orders of magnitude.Recognizing these fundamental trade-offs [12], storage systems like Amazon's SimpleDB and DynamoDB [43], Google's AppEngine data store [22], Yahoo!'s PNUTS [17], and Oracle's NoSQL Database [32] now offer read operations with a choice between strong consistency and eventual consistency. Applications can choose between stronger consistency with lower performance and relaxed consistency with higher performance.A major problem with existing multi-consistency storage services is that application developers need to decide at development time which consistency to embrace. They use different API calls t...

Quantitative Analysis of Consistency in NoSQL Key-Value Stores

“…Thus a natural question derived from those observations is "what consistency does your key-value store provide in practice?". We summarize below the prevailing consistency guarantees that have received considerable attention in recent research on distributed data stores [37,38,28,4]. We will focus on five of them (strong consistency, read your writes, monotonic reads, consistent prefix and causal consistency) in the rest of this paper.…”

“…Note that c ri and c wj are not necessarily different, and the key pairs (e.g., k ri and k rj ) are not necessarily the same; We introduce two notions, causality order and serialization [2,38], before defining S satisfying CC. We say …”

“…It models the main components of Cassandra and its environment such as strategies for ordering multiple versions of data and message delay. We have also specified five consistency guarantees that are largely used in industry, strong consistency (the strongest consistency guarantee), causal consistency (the strongest consistency guarantee that is available in the presence of partitions), read your writes, monotonic reads and consistency prefix (popular intermediate consistency guarantees) [37,38,28,4], in the QuaTEx probabilistic temporal logic [1]. Using the PVeStA [3] statistical model checking tool, we have then quantified the satisfaction of such consistency properties in Cassandra under various conditions such as consistency level combination and operation issuing time latency.…”

“…First, we explain various consistency models in greater detail including their formal definitions (Section 3) and specifications (Section 5.1). Second, we provide quantitative analysis of consistency using both statistical model checking and implementation-based estimation for three new consistency models (Section 5.2): monotonic reads, consistent pefix and causal consistency [37,38,28,4].…”

Billing system CPU time on individual VM

“…This set of resources corresponds to a Service Level Agreement (SLA) which should be fully provided to customers [1], [2]. On their side, providers are interested in saving resources [21], [22] while guaranteeing customers SLA requirements.…”