dovecot and glusterfs

[List]

On 1/13/15, 6:02 AM, Michael Schwartzkopff wrote:
> Am Dienstag, 13. Januar 2015, 21:40:34 schrieb Nick Edwards:
>> On 1/13/15, Michael Schwartzkopff <ms at sys4.de> wrote:
>>> Hi,
>>>
>>> I did some experiments with dovecot on a glusterfs on the active nodes
>>> without
>>> a director. So I had concurrent access to the files.
>>>
>>> With the help of the available documentation about NFS and fcntl locks I
>>> managed to find out the following:
>>>
>>> With the plain mbox format dovecot seems to apply and to honor the fcntl
>>> locks.  But since this format is not used any more in real setups, it is
>>> useless.
>>>
>>> With mdbox and maildir format I could reliably crash my mail storage just
>>> by
>>>
>>> delivering mails to the both dovecots via LMTP to the same user. In
>>> maildir
>>>
>>> dovecot seems not the set / respect the fnctl locks of the index file.
>>> dotlocks
>>> do not seems to work either with mdbox.
>>>
>>> So I think the only solution os to use a director in a real world setup.
>>> Or
>>> is
>>> there any non-obvious trick that I did not check?
>> Interesting, we use NFSv3 dovecot LDA with maildir, we have at present
>> two dozen front end SMTP servers (using dovecot-lda) and some, hrmm we
>> added a few more over Christmas, so I think about 32  pop3 servers,
>> but with only 4 imap servers incl webmail (IMAP is not heavily used
>> here due to government spy laws) talking to NAS storage server
>> backend, *we do not use director* at all and has never been an issue.
>> Director IIRC solves the problem of IMAP inconsistency, but we never
>> see advantage when we tested, no doubt it solves some fancy setup
>> problem, but since director can not help with pop3, it was not worth
>> the hassle. never had any problems with webmail either, load balancers
>> seem to look after it well
> Yes. NFS has its own locking. I wanted to use plain glusterfs client without
> the detour of NFS. Thanks for your hint.
>
> Mit freundlichen Grüßen,
>
> Michael Schwartzkopff
>

The last time we experimented with Glusterfs (two years ago) the native 
client was actually not able to maintain consistency as well as the NFS 
for a reason that I cannot remember anymore.  We used maildir, and when 
using NFS we were able to deliver about a hundred thousand emails per 
hour and do a couple hundred thousand IMAP and POP3 retrievals per hour 
against a modest four node Gluster cluster with four Dovecot/Postfix 
servers (running in vmware).