On 1.4.2011, at 10.52, Jan-Frode Myklebust wrote:

I've notice the log notices about increasing /proc/sys/fs/inotify/max_user_instances on my servers, and started wondering if inotify works for network/cluster filesystems..

I found this:

http://www.ibm.com/developerworks/forums/thread.jspa?threadID=311194

which says that there's no mechanism for one node to tell another node that a directory changed for GPFS.. And I would assume the same applies for NFS and other network filesystems. (Yes I know, we should use director to make users sticky to one host -- but we're not yet ready for that for lmtp).

Right. But it should still work if the modifications come from the same system. So it's better than nothing.

So, are there any recommendations for what settings one should use for for filesystems where inotify woun't work ? Are there settings for disabling it (without recompiling)?

Why not just keep it enabled? There shouldn't be any down side to it.

And does mailbox_idle_check_interval happen in addition to inotify, so that clients will be notified even with a broken inotify ?

Yes.