Replying to list, as one suggestion does involve the database :-)
On Fri, Jan 9, 2009 at 3:03 AM, xufeng <xuf...@yuanjie.net> wrote: > Hi all, > Recently I am redesigning our website architecture. > I have two servers located at Location_A and Location_B serving the same > http://www.ourexample.com and loadbalancing via DNS round-robin.Location_A > and Location_B are connected via Internet(NOT in A LAN). > The mysql database can be "synchronized" via master-master replication as > far as we have seen,but we are now faced with file synchronization > problem(I > mean the files uploaded by users from Location_A and Location_B cannot see > each other "immediately".) Your options depend mainly on speed and reliability between A and B, but are mostly what you already indicated; comments below. > 1.NFS over Internet for file sharing Works, but keep in mind that you'll have one server and one client. This works, but will be relatively slow for the client, and will not provide failover. You *could* probably build half a master-master setup using UnionFS, and then provide close-to-failover using scheduled synchronisations. > 2.sshfs Roughly the same comments as NFS, but on the plus side, more secure and on the minus side, slower > 3.inotify(our system's kernel does not support this and we donot want to > risk upgrading our kernel as well) This would still require you do do some important scripting, anyway. > 4.drbd in active-active mode For drbd to work well in active mode, you still need a networked filesystem, that is, an FS that is aware of modifications on the other side - remember that metadata is cached in the local VFS layer. If you're willing to pay, you could have a look at GPFS, PeerFS or any other number of cluster FSes. These will immediately sync metadata, if not necessarily the actual data. > 5 or any other solutions Well, yes, but only if your file needs remain reasonable: you could ostensibly dump the files in BLOBs in your database, thus leveraging the already existing synchronisation. I strongly recommend you to then also implement a local FS cache layer for them, so that you only need to fetch them from DB once. As said before, though, "immediate" availability strongly depends on the quality of the link between your sites. The advantage of a clustered filesystem, is that it can guarantee that the file is available everywhere as soon as it's been comitted to disk (although the actual first fetch might still be slow). I hope this helps you a little, Johan -- Celsius is based on water temperature. Fahrenheit is based on alcohol temperature. Ergo, Fahrenheit is better than Celsius. QED.
