Re: [Dovecot] Maildir unreliability
On 25.10.2004, at 11:04, Peter Evans wrote:
Are you sure its maildir that is the issue, I am sure God Emperor DJB would have issues with any such statement maligning the holy grail.
The issue looks to me like it is attribute caching, and thus not actually specific to Maildir, just aggravated by it.
Well, it's not maildir issue itself, but maildir more or less relies on readdir() not losing renamed files. I'm not sure what you mean by attribute caching, because it's not just with NFS.
I'm sure if you point that out to DJB he just says it's not a real problem, because it doesn't happen hardly ever except with some stress test tools, or mailboxes with _many_ simultaneous users. Or maybe with a heavily loaded server. Or ...
In any case, I'd rather make Dovecot resistant to this problem entirely rather than hope that everyone uses maildirs only in optimal conditions.
On Mon, 2004-10-25 at 06:44, Timo Sirainen wrote:
On 25.10.2004, at 11:04, Peter Evans wrote:
Are you sure its maildir that is the issue, I am sure God Emperor DJB would have issues with any such statement maligning the holy grail.
The issue looks to me like it is attribute caching, and thus not actually specific to Maildir, just aggravated by it.
Well, it's not maildir issue itself, but maildir more or less relies on readdir() not losing renamed files. I'm not sure what you mean by attribute caching, because it's not just with NFS.
I'm sure if you point that out to DJB he just says it's not a real problem, because it doesn't happen hardly ever except with some stress test tools, or mailboxes with _many_ simultaneous users. Or maybe with a heavily loaded server. Or ...
In any case, I'd rather make Dovecot resistant to this problem entirely rather than hope that everyone uses maildirs only in optimal conditions.
Suggest looking over procmail's empirical approach to optimizing NFS locking.
Ideal might be to have the locking style cached in some file under $PREFIX/var, make dovecot inspect that file when deciding how to do locking, and have an option to dovecot that will redetermine the best style.
It'd make a superb library, since so many apps need to do this.
-- Dan Stromberg DCS/NACS/UCI strombrg@dcs.nac.uci.edu
participants (2)
-
Dan Stromberg
-
Timo Sirainen