Re: Is this a sensible architecture?

[Aaron Richton <richton@nbcs.rutgers.edu>]

On Wed, 14 Mar 2012, Brett Watson wrote:

> I'm currently planning a shift in our use of LDAP to incorporate mirror 
> mode masters for the sake of high availability. The plan is to hide a 
> mirror mode master pair behind a virtual IP using "sorry server" 
> fail-over, such that the primary mirror server takes 100% of the load 
> when it's up and responsive, with fall-back to the secondary mirror 
> server when the primary is down. In this way, the virtual IP presents a 
> "virtual master" to the outside world, and the plan is for *all* outside 
> LDAP interaction with the masters to happen over this single virtual IP.
>
> There will be other "slave" servers which replicate the master (to 
> distribute read-only load), and the plan is for them to syncrepl from 
> the master virtual IP. I gather that a possible alternative to this 
> arrangement is to have the "slave" servers act as syncrepl consumers to 
> *both* masters simultaneously, via their real IP addresses. If this is 
> indeed a valid configuration, does it convey any advantages? The single 
> "virtual master" approach seems architecturally simpler, but is it 
> considered robust by those in the know?   Thanks in advance.

Well, people test this quite regularly (via OpenLDAP's included test 
suite) and it seems to work without issue; there are also live sites using 
mirrormode and (anecdotally) I don't believe they have an excessive defect 
rate versus other installations.

As with most the-other-node-must-be-dead scenarios, the ugly edge cases 
rest more on the traffic direction methods than the actual application. 
You'd probably be well served by taking a careful stock of your "virtual 
IP" methods -- split brain, false positives, what have you -- and OpenLDAP 
will likely serve you well.