Problem: Distributed storage.. in particular, making it easy to program and making it possible for legacy applications to use distributed storage systems.
Is it real? Definitely.
Nugget: WheelFS looks like a standard POSIX file system, but allows specification of distributed storage policies like consistency, primary replica, hints about read patterns, etc to be specified as "cues" in path names.
Will it be influential? The idea of putting hints about consistency/etc properties into "cues" in a path is neat and can probably benefit other systems. Using FUSE to achieve a POSIX-like file system interface and allow legacy apps to run on top is a win; Hadoop's HDFS does this too (sort of) as of recently, and I imagine that other people writing distributed FS'es will do it. The main question with respect to influence of this work will be whether WheelFS will prove flexible and efficient enough for applications beyond the ones it was evaluated with.
Criticism: One general question with a general system like WheelFS is whether a single system can be optimal for all applications. Common engineering sense says that the answer is probably no (although maybe the *API* of WheelFS could be used for other apps). The next question then is, what are the limitations of WheelFS? Can we say which apps it will be good for and which apps it won't?
No comments:
Post a Comment