En/na Jan Wieck ha escrit:

> On 7/1/2004 8:23 AM, Ramon Salvad? wrote:
>
>> -----BEGIN PGP SIGNED MESSAGE-----
>> Hash: SHA1
>>  
>> Hello,
>>
>> ~    I am currently learning how slony works ( I have a master-slave
>> cluster working as intended ). I would like to know the following if
>> possible (or some advice about it):
>>
>> ~    1) Is it wise to use slony to replicate a database of about 80GB
>> (1 Master and 2 or 3 slaves) (about 100 application users concurrently).
>
>
> The size of the database isn't the determining factor. What counts is 
> the read/write ratio of your application. Slony has for example no 
> problem to replicate even the Order-Mix of a TPC-W benchmark, which is 
> about 10-15% write queries with defined think times on the client side 
> (like a real web application). But it has little chance to keep up 
> with a pgbench, which is 75% write queries done by libpq clients 
> hammering transaction after transaction as fast as they can (not 
> really any kind of realistic application).
>
Right, I should have figured that myself. However a sentence in the 
documentacion about the inability to replicate large objects got me a 
bit confused.

>>
>> ~    2) Would it be possible to use LVS to access db slaves as a
>> cluster from the client application (web application)? Is it a good 
>> idea?
>
>
> I would think that depends on the application, and I don't know 
> exactly how LVS works. Each replica is independant, so they can 
> possibly all represent a different point in time. That means, if a 
> user gets a search result from one server, and then tries to do a 
> detail view and the load balancer sends this request to another 
> server, this detail might not have been propagated to this server yet.
>
> If the load balancer honors session ID's and directs all requests 
> belonging to one session to the same server, then there should not be 
> any problem.
>
>
> Jan
>
A LVS Director is a Layer 4 switch so no luck but it supports persistent 
connections (read connection affinity), maybe that should be enough, I 
will try...

Is there any way to keep the slaves in sync? as I understand even if the 
slaves are all using the same source (master) for replication,  it won't 
work.

Thanks
Ramon