Traditional local-area distributed file systems do not meet our goals of speed, availability, and network economy. Systems such as xFS [2] and Frangipani [33] rely on tight node coordination for replica management and cannot overcome the non-uniform networking latencies and frequent network partitioning that are typical in wide-area networks.
Pervasive replication resembles the persistent caching used in client-server file systems such as AFS [13], Coda [20], and LBFS [21]. Pangaea, however, can harness nodes to improve the system's robustness and efficiency. First, it provides better availability. When a server crashes, there are always other nodes providing access to the files it hosted. Updates can be propagated to all live replicas even when some of the servers are unavailable. The decentralized nature of Pangaea also allows any node to be removed (even permanently) transparently to users. Second, Pangaea improves efficiency by propagating updates between nearby nodes, rather than between a client and a fixed server and, creating new replicas from a nearby existing replica. In this sense, Pangaea generalizes the idea of Fluid replication [16] that utilizes surrogate Coda servers placed in strategic (but fixed) locations to improve the performance and availability of the system.
Pangaea's replication follows an optimistic approach similar to that of mobile data-sharing services, such as Lotus Notes [15], TSAE [10], Bayou [32], and Roam [25]. These systems lack replica location management and rely on polling, usually by humans, to discover and exchange updates between replicas. Pangaea keeps track of replicas automatically and distributes updates proactively and transparently to all the users. Most of these systems replicate at the granularity of the whole database (except Roam, which supports subset replicas). In contrast, Pangaea's files and directories are replicated independently, and some of its operations (e.g., ``rename'') affect multiple files, each replicated on a different set of nodes. Such operations demand a new protocol for ensuring consistent outcome after conflicts, as we discuss in Section 5.2. Pangaea offers a simple conflict resolution policy similar to that of Roam, Locus [36], or Coda [18]. We chose this design over more sophisticated approaches (as in Bayou), because Pangaea can make no assumptions about the semantics of file-system operations.
FARSITE [1] and Pangaea both build a unified file system across a federation of nodes, but they have different objectives. FARSITE's goal is to build a reliable service on top of untrusted nodes using Byzantine consensus protocols, and it is designed primarily for local-area networks. Pangaea assumes trusted servers, but it dynamically replicates files at the edge to minimize the use of wide-area networks.
Recent peer-to-peer data sharing systems, built on top of load-balanced, fault-tolerant distributed hash tables, share many properties with Pangaea. Systems such as CFS [6] and PAST [27] employ heuristics to exploit physical proximity when locating data, but they do not support concurrent in-place updates of hierarchically structured data. Pangaea, unlike these systems, provides extra machinery for conflict detection and resolution, as we discuss in Section 5.2. Oceanstore [17] builds a file system with strong consistency by routing updates through a small ``core'' of replicas. Pangaea, instead, allows in-place updating of any replica without centralized coordination to maximize availability. Ivy [22] is a peer-to-peer file system that lets data be updated, any time, anywhere, by exchanging operation logs between replicas. Because the replicas poll remote logs frequently, it supports stronger consistency than Pangaea. Its log-based update propagation also allows for more versatile conflict resolution than in Pangaea. However, because Ivy forces each user to read the logs of all other writers, it can only support a small file system with a small number of writers.
A number of companies are active in the field of wide-area collaborative data sharing, including FileFish, Scale8, WebFS, and Xythos. They offer a uniform, seamless interface for sharing files in a wide-area network, independent of the physical locations of users and data. Some of them provide features such as intelligent location of the cached copy closest to the user. However, they all use a centralized database to keep track of the location of files and replicas. Thus, their design does not meet Pangaea's goals of availability and autonomy.