Re: [389-devel] Thought excersize: A different take on replication

[Gerrard Geldenhuis]

> -Original Message-
> From: 389-devel-boun...@lists.fedoraproject.org [mailto:389-devel-
> boun...@lists.fedoraproject.org] On Behalf Of Soeren Malchow (MCon)
> Sent: 03 December 2010 16:42
> To: 389 Directory server developer discussion.
> Subject: Re: [389-devel] Thought excersize: A different take on replication
> 
> Hi Gerrard,
> 
> Since you are already mentioning OSPF, how about grouping the
> Masterservers into "zones" and have only one of the hosts taking care of the
> communication to other "zones".
> 
> Changes could always only be replicated via the "communication master".
> 
> The main problem that arises here is the potential outage of one of those
> "communication masters", this could be solved by either an election
> process or by an explicit order.

Thanks for the reply Soeren, would this "election" process be on a OSPF 
basis/level? How do you differentiate between the host not available from a 
network point of view and from a service point of view. My network skills are 
limited and to be honest; OSPF was mentioned during our discussions but I had 
to read up on it afterwards and it seemed similar to the solutions we were 
discussing.

> 
> I think this would be a good way to go because it is very close to real life
> scenarios form my point of view, I most likely have a view masters in
> geographically distributed locations, but not tens of servers in one location 
> (
> maybe already counting e.g. different buildings on a campus as locations )
> 
> Not sure if that is a way to go, but maybe it helps a little
> 
Regards


In order to protect our email recipients, Betfair Group use SkyScan from 
MessageLabs to scan all Incoming and Outgoing mail for viruses.


--
389-devel mailing list
389-de...@lists.fedoraproject.org
https://admin.fedoraproject.org/mailman/listinfo/389-devel

[389-devel] Thought excersize: A different take on replication

[Gerrard Geldenhuis]

Hi
We were discussing different setups of replication agreements (multi master) 
between a large number of hosts and ways to minimize contention during updates 
with interconnected hosts. For example the same change might arrive on a host 
from two other hosts via different paths at the same time causing "errors" in 
the log because of exponential back off. If you have to many connections you 
get a replication storm, to little connections and replication takes to long.

The problem to us sounds very much like a network problem or maybe the 
effectiveness of the underlying database to lock the data more effectively. 

We dreamt up a couple of solutions/ideas and I am writing this email to illicit 
some more discussions and/or comments. One solution would be to change the 
underlying database to one that supports improved granular locking (firebird 
comes to mind ) .

Another idea we discussed was based on the following question:
What if you could only define the list of master servers and let the master 
servers figure out the details with regards to doing multi mastering and 
distributing the data and taking care of broken paths? There is similarities 
with OSPF...

Do you have any thoughts on this? Have you had similar ideas? Are we missing 
the point?

Regards


In order to protect our email recipients, Betfair Group use SkyScan from 
MessageLabs to scan all Incoming and Outgoing mail for viruses.


--
389-devel mailing list
389-de...@lists.fedoraproject.org
https://admin.fedoraproject.org/mailman/listinfo/389-devel