Check out the new USENIX Web site. next up previous
Next: Related work Up: Introduction Previous: Graph-based replica management

Optimistic replica coordination

A distributed service faces two inherently conflicting challenges: high availability and strong data consistency [8,37]. Pangaea aims at maximizing availability: at any time, users must be able to read and write any replica and the system must be able to create or remove replicas without blocking.

To address this challenge, Pangaea uses two techniques for replica management. First, it pushes updates to replicas rather than invalidating them, since the former achieves higher availability in a wide area by keeping up-to-date data in more locations. This approach may result in managing unnecessary replicas, wasting both storage space and networking bandwidth. To ameliorate this problem, Pangaea lets each node remove inactive replicas, as discussed in Section 4.4.

Second, Pangaea manages replica contents optimistically. It lets any node issue updates at any time, propagates them among replicas in the background, and detects and resolves conflicts after they happen. Thus, Pangaea supports only ``eventual'' consistency, guaranteeing that a user sees a change made by another user in some unspecified future time. Recent studies, however, reveal that file systems face very little concurrent write sharing, and that users demand consistency only within a window of minutes [31,35]. Pangaea's actual window of inconsistency is around 5 seconds in a wide area, as we show in Section 7.6. In addition, Pangaea provides an option that synchronously pushes updates to all replicas and gives users confirmation of their update delivery (Section 5.3). We thus believe that Pangaea's consistency semantics are sufficient for the ad-hoc data sharing that Pangaea targets.

Pangaea does not support applications that require strong consistency such as open-close consistency, that use locks, or that synchronize using directory operations (i.e., ``lock files'').

next up previous
Next: Related work Up: Introduction Previous: Graph-based replica management
Yasushi Saito 2002-10-08