My understanding of Amazon's internal SOA usage is that it is geared toward timeout-enforced service response guarantees to mitigate concerns about performance penalty. Specifically, a page render event triggers initiation of asynchronous calls to N service endpoints( I have read as many as 300 per page render event ). The page renderer will wait for a specified time to receive responses from each service, but timeout on those services that do not meet their latency SLA. Page components with dependencies on either error result or timed-out service responses will simply be not rendered or replaced with dummy content.

Employees of Amazon, feel free to correct any holes or errors in my understanding of the web application architecture there.