One can use ‘flush’ utility to force clients to invalidate their cache, then update server-side object and next time client reads data from server it will see the new content.

In particular this works for directories as in your example.

Pierre Phaneuf :

I’d be a little weary of people jumping on POHMELFS so quickly… It’s still at a rather early stage, and among other things, has a very small team of one person (who’s been called crazy by the rest of the kernel hacking populace more than once!)…

This is why i keep repeating that it’s staging-level code, and that this is all to be done in a test environment. :) Part of the reason for examining pohmel in a public forum like this is to create more documentation, since it’s not well documented in the first place at all. As well, the more people read about it, the more people might be inclined to get on board with it, which will help with the small/crazy dev problem. ;)

There’s some other interesting options like btrfs+crfs, which have sizeable teams of well-known smart/sane people behind it (as well as tons of money), and is slightly more advanced (although still in development).

But really, in this day and age, I think that trying to provide POSIX semantics in the form of a “real” mountable filesystem *and* scale well has been accepted as just too ridiculously hard and a general waste of time. You have to develop your application in function of the storage system (no MySQL for you!), yes, but for that price, you get otherworldly performance that can bury even POHMELFS (and sometimes even including things such as automatic replication usable in high latency settings, such as replicating between two data centers).

But for sure, the whole subject is rather exciting, no matter which way you go. :-)