Abstract: A system for fault tolerant multi-stage data replication with relaxed coherency comprises one or more processors and memory coupled to the processors, where the memory stores program instructions executable by the processors to implement a data replication service. In response to an update request directed at a data object, in a first stage of replication, the data replication service is configured to update respective primary replicas at a plurality of primary replica sites for the data object, and to store metadata descriptors at each of the plurality of primary replica sites. The data replication service is configured to determine whether a notification of the update is to be sent to any other replica sites. In response to identifying such other replica sites, an update notification including contents of a metadata descriptor may be sent to the other replica sites in one or more additional replication stages.

Abstract: Described is a technology by which data is efficiently and securely stored in a single instance store. A hash value is computed from data in its unencrypted form. The hash value is used to reference a single instance of that data when stored in its encrypted form. In this manner, duplicate data blocks are detectable independent of their encryption, yet stored in an encrypted form in a single instance store. In one aspect, context information for decrypting the encrypted data is stored in association with the data. When the client wants to restore the data, the client sends the hash value for that block to the single instance store service. The service returns the block and the associated context information. The client uses the context information to decrypt the block. For example, the context may comprise a key identifier which the client uses to lookup the correct key.

Abstract: Various embodiments of a system and method for performing a file system operation on a specified storage tier are disclosed. Input specifying a file system operation and a pathname may be received. The pathname may include a plurality of components, including one or more directory components and a storage tier component. The pathname may be analyzed to determine a directory specified by the one or more directory components and a storage tier specified by the storage tier component. The file system operation may be performed, where the file system operation operates on the first directory and the first storage tier. In a further embodiment, the components of the pathname may also include a filename component. Thus, the pathname may be further analyzed to determine a filename specified by the filename component. The file system operation may operate on the first filename, the first directory, and the first storage tier.

Abstract: A plurality of data objects may be replicated across a plurality of computing nodes coupled to a network. The network may include a first node operable to initiate an update operation to update a plurality of replicas of a first object. If one or more of the replicas are not reachable then the update operation may update a subset (e.g., a quorum) but not all of the replicas. For each node on which one of the replicas was updated in the update operation, the node may add the object to a list of incoherent objects. The list of incoherent objects may subsequently be used to bring the lagging replicas in sync with the replicas that were updated. In another embodiment, a plurality of replicas of an object may be stored on a plurality of nodes, similarly as described above. A first node that stores a replica of the object may store a first timestamp associated with the replica on the first node.

Abstract: A system and method for pre-allocating replicas for a distributed file sharing system. Creating a new file may involve creating a plurality of replicas for the file on a plurality of nodes. In one embodiment nodes in the system may pre-allocate sets of file replicas, where the pre-allocated replicas can be used to satisfy requests to create new files. Pre-allocating the file replicas may decrease the latency of file creation requests by enabling the requests to be satisfied without performing replica allocation in response to the requests.