Came back to this account after fifteen years. Half of what I was sure about then is still right. The other half is about Kubernetes now.
The half that survived: architecture decides what scales. EBS is still slower than ephemeral when you’re load-bound. A nine still costs a zero. Sharding is still two steps, and the second one is still the same.
The half that didn’t: the failure modes moved. They didn’t go away. Now it’s the control plane that drops out at 11am, not the disk. The on-call shape changes; the on-call doesn’t.
Same job, new acronyms.