On Tue, 2004-06-01 at 11:00, Brian Candler wrote:
The atomic tmp -> new move prevents of course the case of multiple (inbound, SMTP) writers. Alas following a recent conversation in this ML with Timo all that saves your butt and mail integrity in the case when a large message being read by slow client gets deleted by another client (dovecot instance) is Saint BufferCache. ;)
I don't understand that comment either. If process A has a file open, and process B deletes it, the file remains (in its entirety) on the filesystem until process A closes it. That's not buffer caching; that's the semantics of unlink().
I think Christian was talking about NFS, it doesn't follow the "semantics of unlink()". Rather if a file is deleted and it's tried to be read later you'll get ESTALE.