On Wed, Aug 31, 2022 at 1:03 PM wangchuanlei <wangchuan...@inspur.com> wrote: > > As known to us, if load balancer is used on router, it is centralized. > If the centralized node loses connection to the compute nodes, > the vip is still on the centralized node, at this time, the vip should > not forward packets to backends which on the compute nodes, > right ? >
The job of the health check in the ovn-controller is to see if the backend service is really up or not. The scenario you mentioned is different. In your case it's not the service which is down but the communication between the nodes is lost. OVN enables BFD tunnel on the geneve tunnels to provide HA for the gateway ports but not for the load balancer VIPs. Perhaps you can use the "Forwarding Group" feature of OVN for your use case ? I'm not sure. Check this commit [1] and the relevant documentation/test cases on how to use this feature. If liveness is enabled on a forwarding group, then OVN uses "select" to load balance between the ports of the forwarding group and it also uses "watch_port". Maybe we can add the support of "watch_port" for the OVN load balancers. This needs some investigation. Please feel free to explore this in OVN if you've the bandwidth. Thanks Numan Check out this commit > Thanks > wangchuanlei > > > On Thu, Aug 25, 2022 at 9:54 PM wangchuanlei <wangchuan...@inspur.com> wrote: > > > > Hi, > > For ovn load-balancer(lb), health check is check the status between > > vip and backends. Packet is sent by controller to backends, then > > backends sent a reply, to accomplish the process of healhy check. > > > > On my enviroment, i have two nodes,one is center node, and the other > > is compute node, so both of them has a controller, controller sends > > packet to backends which on its own node. Here, if the center lost > > connection to the node,the status of health check still keep online! > > Because the status are stored in south databse of center. > > In this case, if vip forward the request from client to the backend > > which is on the node, the client would not get a reply, so , here is > > bug! > > I'm not sure if this is a bug. If the compute node loses the connection to the > Southbound database, it can't update the health check status. > > > > As the health check is distributed on controller, should change to > > be centralization ? > > I don't think that would help. What if the centralized node which does the > health check loses connection to the compute nodes ? > > Thanks > Numan > > > > > Thanks! > > _______________________________________________ > > discuss mailing list > > disc...@openvswitch.org > > https://mail.openvswitch.org/mailman/listinfo/ovs-discuss > > > > _______________________________________________ > discuss mailing list > disc...@openvswitch.org > https://mail.openvswitch.org/mailman/listinfo/ovs-discuss > _______________________________________________ discuss mailing list disc...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-discuss
