OVN IDLs not initialized for all worker types
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
neutron |
Fix Released
|
Undecided
|
Unassigned |
Bug Description
Since neutron commit "Rely on worker count for HashRing caching" (https:/
I'm working on an implementation of VPNaaS for OVN (https:/
Secondly our proposed implementation of OVN VPN agent will still make use of rabbitmq RPCs to sync the VPN service settings. The controller implements some RPC methods that may also need access to the AgentCache. So also the RpcWorker might need post_fork_
If I did not overlook something it could be fine to go back to run post_fork_
tags: | added: ovn |
tags: | added: neutron-proactive-backport-potential |
The problem with doing post_fork_ initialize everywhere is that those DB connections are super expensive. Each one maintains an in-memory copy of the OVS DBs, and ML2/OVN itself doesn't use RPC workers at all. If there is a way to let other code opt-in to setting up DB connections as opposed to just enabling it generally for unneeded workers, it'd be nice.