> I wonder why someone hasn't already transitioned Twitter to a publish-subscribe architecture internally, and used a bunch of queues to buffer the impact of lo

> I wonder why someone hasn't already transitioned Twitter to a publish-subscribe architecture internally, and used a bunch of queues to buffer the impact of load spikes ( a slow tweet is not the same as a dropped tweet! ). c.f. Starling : http://rubyforge.org/projects/starling/ http://rubyforge.org/projects/starling/ As others have pointed out, it's not the messaging that's hard, it's the ad-hoc services-- friends-of-friends timelines, keyword search/subscriptions, etc. that are hard to scale and are what is bringing twitter down: http://www.hueniverse.com/hueniverse/2008/03/on-scaling-a-mi.html http://www.hueniverse.com/hueniverse/2008/03/on-scaling-a-mi... Which, to jzawodn's point, is actually really similar to what FriendFeed is doing under the hood, and doing it well.