Comment 0 for bug 1591297

Neutron Metadata Agents works roughly like this: The service runs on a Unix Domain Socket. It runs as many threads as cpu_count/2. Each thread has a maximum number of connections that it can accept. The service attempts to talk to Neutron API via RPC but (only on) Kilo and Liberty will fall back to HTTP if necessary.

By default (since Kilo) Neutron has that number set to 4096. Since kilo in OpenStack Ansible the number has been 128.

At high throughput on a large OpenStack Ansible cloud, keeping this value arbitrarily low (1/32nd of it's default value) causes RPC connections from Neutron Metadata Agent to fail and fallback to HTTP. Setting this value to even half of the upstream default (on only 2 of the 5 metadata agents containers) causes this fallback to stop happening.