You have a virtualized platform, on top of which are many pieces like load-balancing, EBS, RDS, the control-plane itself, etc.
You have burstable network connections which by their nature, will have hard limits (you can't burst above 10Gbps on a 10Gbps pipe, for example; even assuming the host machine is connected to a 10Gbps port).
Burstable (meaning quite frankly, over-provisioned) disk and CPU resources.
And if any piece fails, you may well have downtime...
It is always surprising to me, that people feel that layering complexity upon complexity, will result in greater reliability.
You have burstable network connections which by their nature, will have hard limits (you can't burst above 10Gbps on a 10Gbps pipe, for example; even assuming the host machine is connected to a 10Gbps port).
Burstable (meaning quite frankly, over-provisioned) disk and CPU resources.
And if any piece fails, you may well have downtime...
It is always surprising to me, that people feel that layering complexity upon complexity, will result in greater reliability.