dovecot - Nov 2010 - Local node indexes in a cluster backend with GFS2

Hi, all
this days I'm testing a dovecot setup using lvs, director and a cluster
email backend with two nodes using rhel5 and gfs2. In the two nodes of the
email backend I configured mail location this way:

mail_location sdbox:/var/vmail/%d/%3n/%n/sdbox:INDEX=/var/indexes/%d/%3n/%n

/var/vmail is shared clustered filesystem with GFS2 shared by node1 and
node2

/var/indexes is a local filesystem at the node, so each node has his own
/var/indexes stuff on ext3 and raid1 for improving performance, I mean each
node a different /var/indexes of its own.

Now, if the director was using node2 for the email for user1 and I removed
node2 now all the connections for that user are redirected to node1, that
node has its local /var/indexes, when this ocurr I can not see the emails
that in fact are in the mailboxes, I guess should be indexes not current
with the user mails.
Is supposed indexes to be rebuilt automatically when the node switches?? Or
I have to configure the indexes in the shared storage so the two node update
the information concurrently?
I will appreciate your help...
best regards

Am 15.11.2010 07:44, schrieb Aliet Santiesteban
Sifontes:
> Hi, all
> this days I'm testing a dovecot setup using lvs, director and a cluster
> email backend with two nodes using rhel5 and gfs2. In the two nodes of the
> email backend I configured mail location this way:
> 
> mail_location >
sdbox:/var/vmail/%d/%3n/%n/sdbox:INDEX=/var/indexes/%d/%3n/%n
> 
> /var/vmail is shared clustered filesystem with GFS2 shared by node1 and
> node2
> 
> /var/indexes is a local filesystem at the node, so each node has his own
> /var/indexes stuff on ext3 and raid1 for improving performance, I mean each
> node a different /var/indexes of its own.
> 
> Now, if the director was using node2 for the email for user1 and I removed
> node2 now all the connections for that user are redirected to node1, that
> node has its local /var/indexes, when this ocurr I can not see the emails
> that in fact are in the mailboxes, I guess should be indexes not current
> with the user mails.
> Is supposed indexes to be rebuilt automatically when the node switches?? Or
> I have to configure the indexes in the shared storage so the two node
update
> the information concurrently?
> I will appreciate your help...
> best regards
> 
i have indexes on the cluster storage, in users mailbox dirs
i speculate having another index storage which isnt at cluster may lead
into trouble i some cases, but Timo will know exactly i think

-- 
Best Regards

MfG Robert Schetterer

Germany/Munich/Bavaria

On 15.11.2010, at 6.44, Aliet Santiesteban Sifontes wrote:
> mail_location >
sdbox:/var/vmail/%d/%3n/%n/sdbox:INDEX=/var/indexes/%d/%3n/%n
> 
> /var/vmail is shared clustered filesystem with GFS2 shared by node1 and
> node2
> 
> /var/indexes is a local filesystem at the node, so each node has his own
> /var/indexes stuff on ext3 and raid1 for improving performance, I mean each
> node a different /var/indexes of its own.
This is a bad idea. With dbox the message flags are only stored in index files,
so if you lose indexes you lose all message flags. Users won't be happy.

Should I set mmap_disable = yes when storing indexes in a GFS2 shared
filesystem??

2010/11/15 Aliet Santiesteban Sifontes <alietsantiesteban at gmail.com>
> Ok, I will create a LUN also as a shared clustered storage for indexes, any
> consideration to have into account when the indexes are shared by many
> nodes...
> thank you all...
>
> 2010/11/15 Timo Sirainen <tss at iki.fi>
>
> On 15.11.2010, at 6.44, Aliet Santiesteban Sifontes wrote:
>>
>> > mail_location >> >
sdbox:/var/vmail/%d/%3n/%n/sdbox:INDEX=/var/indexes/%d/%3n/%n
>> >
>> > /var/vmail is shared clustered filesystem with GFS2 shared by
node1 and
>> > node2
>> >
>> > /var/indexes is a local filesystem at the node, so each node has
his own
>> > /var/indexes stuff on ext3 and raid1 for improving performance, I
mean
>> each
>> > node a different /var/indexes of its own.
>>
>> This is a bad idea. With dbox the message flags are only stored in
index
>> files, so if you lose indexes you lose all message flags. Users
won't be
>> happy.
>>
>
>