[simmschi]

At a certain scale you start to break many things that you're used to abstract away into a neat little box. This can be anything, from kernel, libraries, network stacks to anything running in the cloud.

"Serverless" is not really an option at a certain scale. It's likely too expensive, but especially a giant risk if you abstract away something that can fail in an interesting way.

I remember we once managed to break the AWS ELB in one AZ and had a nice little call with a dev from Seattle. Another time we managed to saturate EC2 in a zone at a certain time.

You will need people who can dig into this. If you're at Twitter scale you'll probably want whole teams digging into aspects that you take for granted.