Re: Algorithms and Hashing (was: Re: Reread of config without restart)

Then hashing with a consistent hashing algorithm would helps. Seems like a cool feature doesn't it Willy ?

http://people.csail.mit.edu/karger/Talks/Hash/index.htm http://www.akamai.com/dl/technical_publications/ConsistenHashingandRandomTreesDistributedCachingprotocolsforrelievingHotSpotsontheworldwideweb.pdf

I have already used libketama for memcache and it performed super well http://www.lastfm.fr/user/RJ/journal/2007/04/10/392555/

Patrick

On Thu, Mar 13, 2008 at 11:21 AM, Michael Rennt <m.rennt#gmx.net> wrote:
>
> Patrick Viet wrote:
> > On Wed, Mar 12, 2008 at 8:10 PM, Michael Rennt <m.rennt#gmx.net> wrote:
> >
> >> I see that the server identifier is stored inside the cookies. This is pretty
> >> good, as it keeps stickyness even when there's a failover to another equally
> >> configured loadbalancer.
> >
> >> The only question left for now: How is all the hashing and distribution implemented?
> >>
> >> I see that the source IP (or whatever other parameter) is hashed and divided by
> >> the total weight of all servers. But what if a new server is added?
> >> Won't the stickiness be completely different after that, messing up any existing
> >> stickiness?
> >
> > The hashing choice is done only if there isn't already a persistency cookie.
> > Therefore the old sessions don't move, and the new sessions will
> > benefit from the extra server.
> > You should really try the software. Nothing beats experience.
>
> This is what I did yesterday before sending the mail. I just didn't test source
> persistence together with cookie persistence.
>
> I see that y'all rather use the source option in combination with a persistency
> cookie.
>
> Let me explain why I was asking this question: We are running a different
> software for a customer for 4 years now in a cluster with 100+ machines. On some
> VIPs he's using source persistence only, because cookies are not applicable.
>
> This is the background why I'm afraid that a new server will mess up the stickyness.
>
> Thanks,
>
> Michael
>
>
Received on 2008/03/13 11:27