RE: Failover ordering

[Guus.Leeuw@t-online.de (Dipl.-Inf. Guus Leeuw jr.)]

Gerry,

I've seen this bundling of IP Addresses behind one hostname
at work at a large corporation. It is okay to balance the load,
since all three IP Addresses have a fair chance to get returned
upon request. (Set the cache time in the DNS to 1 second, though,
since local caching these IP Addresses will kill your load balance.)

It becomes very annoying as soon as either master or slave go down.
Following scenarios: Master goes, and a Slave goes down, all Slaves
go down.

Master goes down
================

Nasty, since global updates are not working anymore. This is not
really harmful to your data, though.

If the master is part of the IP Ring, possibly up to 1/3 of your
requests will fail to connect, and hence give weird errors.
(Provided your DNS is not dynamic, and is not updated as soon as
one of the servers go down...)

a Slave goes down
=================

Weird things happen to the queries if either slave go down: Some
of your requests will return "Connection Error", while others still
work. If replication is off (as in iPlanet 4.x and before), the slave
that should receive backlogs as soon as it is up again, might invalidate
the (his) data, giving troubles when UNIX/Windows accounts are linked to
userPassword, passwordExpirationTime and the likes. Also, your accountant
data mught be corrupt. I've seen a situation where even cn, or sn from
one slave to another were out of sync under Netscape / iPlanet.

all Slaves go down
==================
If the master is not in the IP Ring, nobody will be able to read data...
(Master not in the Ring, means master isn't used for reading, to me at
least.)
If the master is in the Ring, it might have a hard time dealing with all
the requests.
Same problems as above might occur.

Based on my inexperience with replication under OpenLDAP, I cannot say
whether replication might wander out of sync with this implementation as
well.

As far as your fail over is concerned, I cannot say anything, since I've
never seen a fail over saved LDAP Server. (Kinda hard, as well, I imagine,
since what would fail over? And who has the right data? Where to store it?
And what if those disks are corrupted... Better chance against disk
corruption
is replication to a different area, but that holds the pitfalls mentioned
above...)

Problematic :) More insights?

Guus
>  -----Original Message-----
>  From: owner-openldap-software@OpenLDAP.org
>  [mailto:owner-openldap-software@OpenLDAP.org]On Behalf Of
>  Morong, Gerry
>  Sent: Monday, July 16, 2001 10:38 PM
>  To: 'openldap-software@OpenLDAP.org'
>  Subject: Failover ordering
>
>
>
>  We have one master and two replica LDAP servers in a test
>  mode.  What is the
>  better way to configure that /etc/ldap.conf HOST line for
>  failover and
>  balanced work load?  If you list all three of the LDAP
>  server on the HOST
>  line, do they failover sequentially or is a server chosen
>  randomly from that
>  line?  Is it better to configure DNS with one hostname that
>  points to all
>  three server IP address?  Would this kill your failover?  Any
>  recommendations??
>
>
>  Gerry
>