Monday, April 20, 2009


Problem: There is a growing set of Internet-based services that are too big, or too important, to run at a single site. Examples include Web services for e-mail, video and image hosting, and social networking. This is the same reasoning behind Facebook’s need to scale out (see below).

Solution: WheelFS is a wide-area distributed storage system intended to help multi-site applications share data and gain fault tolerance. Different distributed applications might need different properties in a storage system: they might need to see the latest copy of some data, and be willing to pay a price in high delay, or they may want data to be stored durably, or have specific preferences for which site stores a document. Thus, in order to be a usable component in many different systems, a distributed storage system needs to expose a level of control to the surrounding application.
WheelFS allows applications to adjust the default semantics and policies with semantic cues, while maintaining the standard POSIX interface. WheelFS offers a small set of cues in four categories: placement, durability, consistency, and large reads.

Tradeoffs: Every file or directory object has a single “primary” WheelFS storage server that is responsible for maintaining the latest contents of that object. WheelFS maintains the POSIX interface even in the background and accessing a single file could result in communication with several servers, since each subdirectory in the path could be served by a different primary.
In addition, a disadvantage of cues is that they may break software that parses pathnames and assumes that a cue is a directory. Another is that links to pathnames that contain cues may trigger unintuitive behavior. However, they have not encountered examples of these problems.

Future Influence: The main insight from WheelFS is that different distributed applications need different properties in a storage system. In order to use have a distributed storage system as a usable component upon which many different applications are built, the storage system must allow the application to control various policies.
It is unclear whether a truly wide-area distributed storage system is the best approach. A middle ground may be found with distributed storage within each datacenter with updates pushed out to the other datacenters. This will create greater isolation between datacenters to prevent cascading failures.

No comments:

Post a Comment