I remember reading a blog post a few years ago where someone tested the Noisy Neighbor theory on AWS, and the results were surprising (Two instances of the same type would have as much as a 50% performance difference in some metrics, Disk I/O being a big one, CPU contention another). I’m surprised that hasn’t been done recently, maybe because hardware upgrades and the workload patterns of most users have made the issue “go away”. I cringe a little inside when I hear the word vCPU thrown around as a unit of performance like that’s supposed to mean anything.