In order to improve scaling and reduce the load to OVSDB SB servers, Neutron could connect to OVSDB replication SB servers in read-only mode. So far, the main use case of writing to SB database is the agent liveness checks where Metadata agents write to the Chassis Private table to reflect that they are alive. Theoretically, this can be avoided by writing into local OVSDB and let ovn-controller copy it over to the SB database. The purpose of this BZ is to research if we can stop writing (or at least reduce the number of 'writing' connections) to SB database and distribute the reading load across the Active/Passive cluster. Thanks to Ilya Maximets & Dumitru Ceara for the ideas.
(In reply to Daniel Alvarez Sanchez from comment #0) > In order to improve scaling and reduce the load to OVSDB SB servers, Neutron > could connect to OVSDB replication SB servers in read-only mode. > > So far, the main use case of writing to SB database is the agent liveness > checks where Metadata agents write to the Chassis Private table to reflect > that they are alive. Theoretically, this can be avoided by writing into > local OVSDB and let ovn-controller copy it over to the SB database. > There´s a drawback with this approach. The Chassis_Private table was introduced to avoid triggering events every where when we write to the Chassis table and currently, updating the local OVSDB will make ovn-controller to write to the Chassis table again (not Chassis_Private) which will re-introduce the triggering problem.
I think the agent delete support patch (https://review.opendev.org/c/openstack/neutron/+/752795/) already removes our writes to sbdb.