From: Lionel Bouton <lionel-subscription@...>
Date: 2007-07-18T05:18:18+09:00
Subject: Re: memcache C extension

snacktime wrote the following on 17.07.2007 21:45 :
> I've decided I want to become more proficient in C, so the project is
> chose is a ruby memcache C extension.  I found the following site with
> a hash algorithm that seems to be an improvement on the default on,
> can anyone think of a reason why not to use this?
>
> http://www.last.fm/user/RJ/journal/2007/04/10/392555/

- very poor distribution with few servers,
- no control of weight between servers (ie if I have a memcache server 
with 2GB and one with 128MB, nothing prevents the second from getting 10 
times as much access as the first one...).


This algorithm benefits very large server farms with identical server 
configurations where having poor performance over a small part of the 
cache isn't a big deal.

If you want both good controlled distribution and high cache hits with 
low number of servers, on first thought I'll implement a double query 
system where cache clients are responsible for moving data accross 
servers based on an old configuration and a new one.
When the hit rate ratio on the new configuration is good and bad on the 
old one, deactivate queries on the old configuration.
Assuming 2 gets and a put to memcache servers are faster than direct 
access to the data, this is probably a win.

With that kind of system you could even handle server failures 
automatically by creating new configurations on the fly (and using hit 
ratio stats from the two previous configuration to decide which one is 
the better new 'old one').

But then again it's only a first thought :-)

Lionel.