there's some speculation that twitter's gonna die because the new ceo is cutting cloud spending, but that betrays a lack of understanding of how twitter works
tweets and DMs are in the datacenter, mostly in aurora/mesos. there isn't autoscaling in any recognizable AWS sense.
traffic is failed-over between datacenters. this is why your follower count occasionally flips between 665 and 666 - one datacenter has the wrong count or hasn't processed the unfollow in its backed-up queue, but your request landed there anyway.
what's more likely to happen is that they'll break a circular dependency, accidentally make it so the frontend can't communicate with the graph-store holding the posts, be unable to roll it back, and cause prolonged outages
or, they'll lose power in one datacenter and be unable to get services restarted, consistency checks performed, and caches warmed, before the other datacenter takes a power hit
in short, twitter is more like the windscale piles than chernobyl — it is going to fail by giving extremely normal readings in one datacenter, extremely bad readings in the other, and the operators are going to open it up and find cherry-red graphite that flashes over into an intractable cleanup operation
(nitpickers corner: some features (spaces, periscope) are cloud-backed)
