[Dovecot] Proxy IMAP/POP/ManageSieve/SMTP in a large cluster enviroment

[Stan Hoeppner]

Aliet Santiesteban Sifontes put forth on 7/18/2010 4:22 PM:

> 6- Mail backend.
>  - ( n servers)RHEL5/6 using Dovecot.
> 
> Now for functional role 6 "Mail Backend" we have some dilemmas.
>    - Recommended scalable filesystem to use for such scenario(Clustered or
> not).

You have 3 choices:  NFS, a cluster filesystem, or each of those with VMWare
ESX.  I recommend the cluster filesystem solution, specifically SGI CXFS, atop
ESX (if you can afford it):

http://www.sgi.com/pdfs/2816.pdf
http://www.vmware.com/products/vsphere/esxi-and-esx/index.html

>      -GFS2?? We have very bad experiences with GFS1 and maildir, GFS2
> doesn't seems to improve this also. Using  some techniques for session
> afinity with a backend server seems to help with the locking problems of GFS
> and the cache. Using GFS many IMAP SMTP servers can in parallel write or
> read to user mailboxes, if GFS can perform well we prefer this, will have to
> proxy o use director for the cache problem.

CXFS doesn't suffer these problems associated with GFS1/2.

>      - Ext4, Ext3 well tested with many setups but only one backend server
> can access one ext4 lun, so we have to proxy a user always to a ip backend
> address or VIP in a case of failure cluster software has to move the server
> to another node, the cluster service is formed by a IP Address and one or
> many Luns wich moves from node to node in a case of failure.

Again, CXFS eliminates this problem as all clients "own" all LUNs all the
time.  Or are you talking about booting the cluster member machines from SAN
LUNs, i.e. "Boot from SAN", /boot and / reside on a SAN LUN?  In that case,
using VMWare ESX server with High Availability and Vmotion to move or restart
failed cluster guests would be a much better more manageable solution as each
guests' system disk would reside within a VMWare virtual disk which would
physically be a single SAN LUN.  You could stick 100 or 1000 virtual machine
disk images on that single LUN, each of them being the boot disk for one of
your mail cluster hosts.

>   - Recommended scalable method  for partition divition.
>      - Directory Hashing(How it actually works the theory behind this), we
> see the wiki but we need to understand the theory to balance the directory
> structure.
>      - Using some criteria ex Lun1(a-h users), Lun2(i-m users)....

You're talking about the mail store directory structure for the IMAP servers
here right?  I wouldn't do the split you describe above.  It will cause more
problems than it will solve.  Again, you can partition and format a single SAN
LUN, say 10TB, with CXFS and have all back end IMAP server hosts mount this
same filesystem.  You can put well over 1 million directories and files within
this CXFS filesystem without seeing performance degradation due to CXFS overhead.

>  - Proxy or Director is a must for a Clustered Filsystem or for the solution
> of a user belonging to one server.
>      - In case we use a Proxy with dovecot we know we can use it with
> IMAP/POP but are not sure for ManageSieve and more important for SMTP. The

You should try to proxy all IMAP/POP client connections to the same Dovecot
server, not necessarily from a data integrity/file locking standpoint (if
using CXFS locking issues are moot), but more from a performance and
efficiency standpoint due to Dovecot indexing.

Regarding inbound SMTP, you don't need to worry about locking here either.
You will have both, or 4, or however many MX hosts connected to the SAN and
all of them will run the CXFS client, Postfix and Dovecot/LDA/ManageSieve.
They will all mount the same SAN LUN that the backend IMAP servers do, and
Postfix/Dovecot LDA will drop the mail into the appropriate user directory
after Sieve scripts are run.  You will create a directory within the user mail
storage LUN where you will store ManageSieve scripts.  This allows for all
inbound SMTP servers to read/write the directory where the ManageSieve scripts
are kept.  Thus, you'll run ManageSieve on all your MX hosts.  It won't matter
if you proxy MS operations or not because they are so infrequent and you'll
not have any locking issues.  But by running MS on all the MX hosts you enable
redundancy.  The only thing you need a director for in this case is routing
requests around a downed MX host.

> question is, how can we proxy for the same user IMAP/POP/ManageSieve and
> SMTP delivery to mailbox at the same time, can the incoming mail gateways
> send the email to the proxy servers and dovecot proxy the request for smtp
> or lmtp appply the same proxy criteria that uses for the protocols IMAP and
> POP and send the email to the correct backend server and this server do the
> final delivery?? I mean proxy all the protocols for a user
> IMAP/POP/SMTP-LMTP?? Any example...

No.  See above, and, IMAP/POP requests would be proxied to the backend IMAP
servers.  Inbound SMTP and ManageSieve requests should be routed to the MX
hosts, no need to proxy or balance them with any affinity.  This is because
SMTP/ManageSieve packets aren't stateful in nature like IMAP/POP are.

> Right now we have doubts on this stuff, we are really going to appreciate
> any help or advice on this...
> best regards

I'm not an expert on all this stuff, and I may not be 100% accurate with all
my advice.  I'm much more knowledgeable on the fiber channel SAN, filesystem
(XFS/CXFS), and VMware stuff than Dovecot and IMAP clustering specifically.  I
hope I've at least helped point you in a good direction in some small way.

-- 
Stan