Reliability at massive scale is one of the biggest challenges we face at Amazon.com, one of the largest e-commerce operations in the world; even the slightest outage has significant financial consequences and impacts customer trust. The Amazon.com platform, which provides services for many web sites worldwide, is implemented on top of an infrastructure of tens of thousands of servers and network components located in many datacenters around the world. At this scale, small and large components fail continuously and the way persistent state is managed in the face of these failures drives the reliability and scalability of the software systems. This paper presents the design and implementation of Dynamo, a highly available key-value storage system that some of Amazon’s core services use to provide an “always-on ” experience. To achieve this level of availability, Dynamo sacrifices consistency under certain failure scenarios. It makes extensive use of object versioning and application-assisted conflict resolution in a manner that provides a novel interface for developers to use.

Peer-to-peer (P2P) online communities are commonly perceived as an environment offering both opportunities and threats. One way to

This paper presents a decentralized, peer-to-peer web cache called Squirrel. The key idea is to enable web browsers on desktop machines to share their local caches, to form an efficient and scalable web cache, without the need for dedicated hardware and the associated administrative cost. We propose and evaluate decentralized web caching algorithms for Squirrel, and discover that it exhibits performance comparable to a centralized web cache in terms of hit ratio, bandwidth usage and latency. It also achieves the benefits of decentralization, such as being scalable, self-organizing and resilient to node failures, while imposing low overhead on the participating nodes. 1.

We present Symphony, a novel protocol for maintaining distributed hash tables in a wide area network. The key idea is to arrange all participants along a ring and equip them with long distance contacts drawn from a family of harmonic distributions. Through simulation, we demonstrate that our construction is scalable, flexible, stable in the presence of frequent updates and offers small average latency with only a handful of long distance links per node. The cost of updates when hosts join and leave is small.

In this paper, we argue that there is a need for an event-based middleware to build large-scale distributed systems. Existing publish/subscribe systems still have limitations compared to invocation-based middlewares. We introduce Hermes, a novel event-based distributed middleware architecture that follows a type- and attribute-based publish/subscribe model. It centres around the notion of an event type and supports features commonly known from object-oriented languages like type hierarchies and supertype subscriptions. A scalable routing algorithm using an overlay routing network is presented that avoids global broadcasts by creating rendezvous nodes. Fault-tolerance mechanisms that can cope with different kinds' of failures in the middleware are integrated with the routing algorithm resulting in a scalable and robust system.

We propose a new approach for constructing P2P networks based on a dynamic decomposition of a continuous space into cells corresponding to processors. We demonstrate the power of these design rules by suggesting two new architectures, one for DHT (Distributed Hash Table) and the other for dynamic expander networks. The DHT network, which we call Distance Halving, allows logarithmic routing and load, while preserving constant degrees. Our second construction builds a network that is guaranteed to be an expander. The resulting topologies are simple to maintain and implement. Their simplicity makes it easy to modify and add protocols. We show it is possible to reduce the dilation and the load of the DHT with a small increase of the degree. We present a provably good protocol for relieving hot spots and a construction with high fault tolerance. Finally we show that, using our approach, it is possible to construct any family of constant degree graphs in a dynamic environment, though with worst parameters. Therefore we expect that more distributed data structures could be designed and implemented in a dynamic environment.

This paper discusses the feasibility of peer-to-peer full-text keyword search of the Web. Two classes of keyword search techniques are in use or have been proposed: flooding of queries over an overlay network (as in Gnutella), and intersection of index lists stored in a distributed hash table. We present a simple feasibility analysis based on the resource constraints and search workload. Our study suggests that the peer-to-peer network does not have enough capacity to make naive use of either of search techniques attractive for Web search. The paper presents a number of existing and novel optimizations for P2P search based on distributed hash tables, estimates their effects on performance, and concludes that in combination these optimizations would bring the problem to within an order of magnitude of feasibility. The paper suggests a number of compromises that might achieve the last order of magnitude.

Unstructured overlays form an important class of peer-to-peer networks, notably when content-based searching is at stake. The construction of these overlays, which is essentially a membership management issue, is crucial. Ideally, the resulting overlays should have low diameter and be resilient to massive node failures, which are both characteristic properties of random graphs. In addition, they should be able to deal with a high node churn (i.e., expect high-frequency membership changes). Inexpensive membership management while retaining random-graph properties is therefore important. In this paper, we describe a novel gossip-based membership management protocol that meets these requirements. Our protocol is shown to construct graphs that have low diameter, low clustering, highly symmetric node degrees, and that are highly resilient to massive node failures. Moreover, we show that the protocol is highly reactive to restoring randomness when a large number of nodes fail. KEY WORDS: Membership management; peer-to-peer; epidemic/gossiping protocols; unstructured overlays; random graphs.

It has been reported [26] that life holds but two certainties, death and taxes. And indeed, despite much effort devoted to circumventing both phenomena, it does appear that any society—and in the context of this paper, any large-scale distributed system—must address both death (failure) and the establishment and maintenance of infrastructure (which we assert is a major motivation for taxes, so as to

In this paper, we present Chainsaw, a p2p overlay multicast system that completely eliminates trees. Peers are notified of new packets by their neighbors and must explicitly request a packet from a neighbor in order to receive it. This way, duplicate data can be eliminated and a peer can ensure it receives all packets. We show with simulations that Chainsaw has a short startup time, good resilience to catastrophic failure and essentially no packet loss. We support this argument with real-world experiments on Planetlab and compare Chainsaw to Bullet and Splitstream using MACEDON.