Citation Context

...e challenges in achieving strong consistency and high availability in a poorly-connected network environment. Some systems address this by reducing consistency guarantees to achieve high availability =-=[22,14,6]-=-. But this shifts the burden to the applications to deal with conflicting views of data. For instance, Amazon’s SimpleDB was originally introduced with an eventual consistency model and more recently ...

Citation Context

...of this data model, however, makes building a number of interesting services overly arduous, particularly compared to the column-family data models offered by systems like BigTable [19] and Cassandra =-=[37]-=-. These rich data models provide hierarchical sorted column-families and numerical counters. Column-families are well-matched to services such as Facebook, while counter columns are particularly usefu...

Citation Context

...mary ofsthat work. More recently, shared-nothing key-value stores havesbeen developed for cloud computing with similar goals, such assGoogle’s Bigtable [8], Amazon’s Dynamo [15], Facebook’s Cassandra =-=[21]-=-, and Yahoo!’s PNUTS [10]. However, unlike Hyder,sthese systems do not support ACID transactions and they requiresthe database to be partitioned. Microsoft’s SQL Azure [3][6] andsElasTraS [14] are clo...

Citation Context

...ies lock contention as a major bottleneck for multicore databases, and improves performance by removing locks incrementally. Masstree provides high concurrency from the start. Recent key-value stores =-=[2, 3, 12, 17, 26]-=- provide high performance partially by offering a simpler query and data model than relational databases, and partially by partitioning data over a cluster of servers. Masstree adopts the first idea. ...

Citation Context

...fically takes advantage of the transience of the data it stores to improve space e ciency. Recent distributed key-value systems such as Google’s Bigtable [5], Amazon’s Dynamo [10], Apache’s Cassandra =-=[17]-=-, FAWN-KV [2] and Ramcloud [24] aimed primarily for high availability, scalability and load balancing among a large number of storage nodes. Our system focuses on boosting the performance on a single ...

Citation Context

...ter tweets or Facebook wall updates) and other users reading them. Both workloads execute 2M operations on values consisting of 10 columns with 100 byte fields. We compare Tembo to Cassandra (v0.6.1) =-=[29]-=-, a distributed data store that borrows concepts from BigTable [10] and Dynamo [16]. We used three different Cassandra configurations in this experiment. The 11Throughput (Ops/sec) 200000 150000 1000...

Citation Context

...w generation of distributed data platforms is the adoption of a range of weak consistency models, such as eventual consistency [1], restricted transactional semantics (e.g. single object transactions =-=[2]-=-, or static transactions whose accessed dataset is pre-declared [3]), and non-serializable isolation levels [4]. The point is that classical strongly consistent transactional replication protocols, ba...

Citation Context

...onsistency, Performance 1. INTRODUCTION Modern distributed applications are reshaping the landscape of storage systems. Recently emerging distributed key-value stores such as BigTable [11], Cassandra =-=[32]-=- and Dynamo [19] form the backbone of large commercial applications because they offer scalability and availability properties that traditional database systems simply cannot provide. Yet these proper...

Citation Context

...s that have high rates of membership churn [13, 42, 48, 51]. Cluster systems, even though they have much lower churn than Internetwide systems, have also used consistent hashing for data partitioning =-=[15, 32]-=-, but have faced interesting challenges. As observed in Amazon’s Dynamo, consistent hashing’s data distribution has a high load variance, even after using “virtual servers” to map multiple randomly si...