Hi,

I posted a roll-up JIRA to address the general problem of network partitions.  
This needs to be fixed.

That said, I don't think it will be easy to fix Sequoia to work over a WAN.  
There are a couple of reasons for this:

1.) Network partitions inevitably lead to re-initializing data on either one 
side or the other side of the WAN.  This is true even if you properly handle 
the partition itself.

2.) WAN connections have very high latency compared to a LAN.  This is a 
problem for group communications, which either pass tokens between members or 
depend on message exchange to implement total ordering.  For example a basic 
LAN with a reasonably capable switch has ping response on the order of .1ms.  
In theory this would allow you to order up to 10,000 messages per second 
assuming a message exchange for each message ordered.  This is the top rate for 
updates.  On the other hand, if you have a 30ms ping, this would imply you can 
order about 30 messages per second.  That limits you to 30 updates per second 
or so.  That's just not acceptable for most applications.

Both of these limitations are inherent in systems that use group communication 
for replication.  This design approach only works on networks that have LAN or 
near LAN availability and latency.  There have been some research papers about 
GC over WAN but they did not look very compelling.

The best way to implement WAN clustering is to use asynchronous replication to 
connect clusters.  This is a very different approach and either requires some 
additional support in Sequoia or integration with database replication such as 
MySQL replication or PostgreSQL SLONY.

Thanks, Robert

P.s., please feel free to correct my math.

On 3/31/08 1:54 AM, "Stefan Lischke" <[EMAIL PROTECTED]> wrote:

Hi Roy,

Use of Sequoia over the internet or even through some not directly
connected network is very unstable.
We are using this scenario but we have a lot of trouble with it, because
sequoia does not handle network partitions. And network partitions occur
very often in such networks.

If there is a network partition, sequoia stops distributing the insert
requests to all nodes, but when the connection is available again,
sequoia distributes alle insert request to all nodes again without
checking the state of the databases on every node. That results in ugly
errors,  because some rows are not present on all nodes.

There is already a Bug for this

https://forge.continuent.org/jira/browse/SEQUOIA-980

And maybe there will be a solution for this in near future.

hth

Stefan

Morris, Roy wrote:
> I was reading through the archives trying to figure if anyone has had 
> controllers
> on multiple sites and it looks like at some point in 2005 this thread ends.
> https://forge.continuent.org/pipermail/sequoia/2005-November/000152.html
>
> I didn't see any real resolution either, I am surprised there are not some 
> huge
> installations out there that require multi-site ability. Anyway if anyone is 
> doing
> multi-site drop me a line I'd like to avoid building a tool for this if I can.
>
> cheers
> Roy
>
>
> _______________________________________________
> Sequoia mailing list
> [email protected]
> https://forge.continuent.org/mailman/listinfo/sequoia
>
>
>


_______________________________________________
Sequoia mailing list
[email protected]
https://forge.continuent.org/mailman/listinfo/sequoia



--
Robert Hodges, CTO, Continuent, Inc.
Email:  [EMAIL PROTECTED]
Mobile:  +1-510-501-3728  Skype:  hodgesrm
_______________________________________________
Sequoia mailing list
[email protected]
https://forge.continuent.org/mailman/listinfo/sequoia

Reply via email to