sharding (sic!) web tier method hinder bucket balancer bottleneck?

how infinite web sites can't totally stateless grasp impassioned scalability during web tier?

there sites ebay amazon, can't totally stateless, have offering transport something that. isn't presumably encode each vigilant offering transport url, nor presumably encode each vigilant cookie send during each connection. jungle simply stores session-id cookie being sent. i know scalability web tier ebay jungle should many harder scalability google hunt engine, where all encoded calm url.

on hand, both ebay good jungle scaled definitely massively. gossip there 15000 j2ee concentration servers during ebay.

how sites hoop both: impassioned scalability statefulness? site stateful, isn't presumably rudimentary dns-balancing. assume companies have hardware formed bucket balancer bigip, netscaler something that, unique device behind unparalleled ip castle site. bucket balancer decrypt ssl (if encoded), check cookie endorse depending event id cookie concentration server binds event customer.

but only can't presumably work unparalleled load-balancer presumably hoop bucket thousands concentration servers? i suspect even hardware bucket balancers scale such level.

also, load-balancing being finished transparently user, i.e. users forwarded opposite addresses, nonetheless still collectively stay during www.amazon.com whole time.

so doubt is: there special pretence grasp something pristine sharding web tier (not database tier finished commonly)? enlarged cookie isn't legalised there proceed know concentration server holding session.

edit: i satisfied there wholly need transparency, there need site spidered bookmarked. e.g. site small web app, something qualification sight sheet reservation system, there should problem only redirecting users specific clusters web servers behind opposite urls, e.g. a17.ticketreservation.com. specific case, presumably only churned clusters concentration servers, any behind possess bucket balancer.
interestingly, i site uses kind concept.
edit: i found visualisation during , where contention refers an essay symbol zhu named . symbol zhu uses querulous scripting fan side bucket balancing transparently.

even there drawbacks, bookmarking, xss, etc, i cruise sounds severely good thought certain special situations, namely roughly content-free web applications, indispensable spidered bookmarked (e.g. sheet reservation systems something that). following there need bucket balancing transparently.

there rudimentary route sure site server, e.g. route www.ticketreservation.com a17.ticketreservation.com. there user stays during server a17. a17 server, nonetheless cluster itself, excess achieved.

the initial route server itself cluster behind bucket balancer. way, unequivocally high scalability achieved, primary bucket balancer behind www wholly strike once during commencement any session.

of course, route opposite urls looks severely nasty, nonetheless small web applications (which need spidered, deep-linked deep-bookmarked anyway), should wholly an visible problem user?

the redirect-cluster check bucket concentration clusters adjust redirects accordingly, so achieving balancing small bucket distribution.