Is this still using the EBS RAID that you guys mentioned in a blog post a while ago? If so, how do you avoid any slow I/O requests (which plague EBS) stalling all I/O to the volume?
Actually, Postgres WAL support helps mitigate this reasonably well, and a well tuned application has something like a cache hit rate of >99% during normal operation.
That said, yes, EBS can have unstable performance. We monitor for several kinds of related problems and in the most dire of straights can perform a hardware migration to a quieter node which generally clears issues up.
