Comment 6 for bug 1245541

About the scheduler filter (sorry for my late relpy), I believe I've used ComputeCapabilitiesFilter.

About the usage:

Actually we've used the flag to specify heterogeneous architecture support. In addition to x86 architecture, we wanted to support other architecture using the instance_type_extra_specs flag.
And we put dynamically changing values in the flag.
For example,

instance_type_extra_specs=cpu_arch:x86_64, gpu_arch:fermi, gpus:2, hypervisor_type:LXC

shows that the nova-compute node has two GPUs of NVdia fermi and runs LXC hypervisor.
The 'gpus' key is dynamically changing at run-time as gpus are allocated and deallocated.
Most static keys can be handled by host-aggregate method.
But keys with dynamic values cannot be handled by host-aggregate method.

Even before the merge of "https://github.com/openstack/nova/commit/4b4f0d61836e527340511f6198c2f2c0bdf70888 merged in 2014-01-28", the dynamic update mechanism was broken.