Nova is built on a shared-nothing, messaging-based architecture. All of the major nova components can be run on multiple servers. This means that most component to component communication must go via message queue. In order to avoid blocking each component while waiting for a response, we use deferred objects, with a callback that gets triggered when a response is received.
In order to achieve shared-nothing with multiple copies of the same component (especially when the component is an API server that needs to reply with state information in a timely fashion), we need to keep all of our system state in a distributed data system. Updates to system state are written into this system, using atomic transactions when necessary. Requests for state are read out of this system. In limited cases, these read calls are memoized within controllers for short periods of time. (Such a limited case would be, for instance, the current list of system users.)
Below you will find a helpful explanation.
[ User Manager ] ---- ( LDAP )
|
| / [ Storage ] - ( ATAoE )
[ API server ] -> [ Cloud ] < AMQP >
| \ [ Nodes ] - ( libvirt/kvm )
< HTTP >
|
[ S3 ]