[Gross] N-way cluster

[Jesse Thompson]

On 04/30/2010 03:25 AM, Alvaro Marín wrote:
> Hi all,
>
> new to this list :)
> Searching for scripts/programs to apply greylisting to dynamic ranges I
> found MARBL, and then I found gross. I've test it and runs fine!
>
> Now, I want to implement it on my production environment, with 13
> antispam servers, but I've read about "problems" with replication with
> clusters with more than 2 servers and this message:
>
> https://lists.utu.fi/pipermail/gross/2008/000142.html
>
> but I suposse that it'll be ready for the v2 of gross, with the new
> architecture:
>
> http://code.google.com/p/gross/wiki/VersionTwoArchitecture
>
> and not for this v1. Am I rigth?

If I interpret Eino's plan correctly, n-node clustering will benefit you 
in terms of scalability, not fault tolerance.  You will still be 
dependent on the 2 master servers being available for the service to 
function.  So, the only reason you would need n-node clustering is if 
your load is so high that 2 nodes can't handle the traffic.  However, if 
the 2 nodes can handle your peak traffic, then you don't actually need 
the n-node cluster.

I'd be willing to bet that 2 nodes could handle your traffic just fine. 
  We had one day where a misconfigured spam bot attempted to send 100 
million messages.  Gross greylisted them all.  I didn't see any adverse 
effects from this load.  I didn't even know about it until I looked at 
our daily message volume report the following day.

Also, keep in mind that you may not need to process as many messages as 
you think.  Greylisting causes a dampening effect on your overall 
volumes.  Before gross we were peaking at 10 million messages/day.  Now 
we are down to a measly 3 million (with only 15% of the messages that 
make it past gross being spam.)


> Another question...I read on the FAQ:
>
> ===
> Q: Are there risks of database problems?
>
> Answer: [...] If you get more than 10 million messages a day you might
> need to adjust the default values.
> ===
>
> I receive ~50mill/day so I've to adjust the default values...should I
> increase these ones?
>
> filter_bits=24
> number_buffers=8

We ran into false matches, and I do remember changing those settings. 
According to the doc: "lowering filter_bits will increase the 
probability of false matches".  We are at 22, which is lower than the 
default of 24.  Perhaps the default value used to be lower than 22?

Or, I think that you could raise number_buffers to decrease the 
probability of false matches.  We do have this setting raised to 12 
(from the default of 8)

I hope this helps!

Jesse

>
> Thank you!
> Regards,
>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 3317 bytes
Desc: S/MIME Cryptographic Signature
URL: <http://lists.utu.fi/pipermail/gross/attachments/20100430/dec97fc1/attachment.p7s>