On 05/17/2012 11:28 PM, Christoph Hellwig wrote: > I startes to look a bit over the object cache, and the async cache > flush option scares me. It only replies to the request before committing > data to the backing storage, thus breaking the transactional sematics > of any filesysem using the volume, as well as removing the possibily > to return errors for the cache flush. > > What is the intended use case for it? I would be much more happy if > this breakage of semantics wasn't offered at all. This is supposed to be an interim solution for the inefficiency of current recovery against heavy IO. Currently, without it, we can't run massive nodes at all, even a single node event will put the cluster into useless. I'd like to remove it until we have settle down all the inefficiency problem in recovery code. Thanks, Yuan |