Bug #62637
openmgr/cephadm haproxy configuration uses the host IPs for Ganesha backend
0%
Description
When Ganesha is deployed, it's highly likely that it will be a multihomed server. This requires haproxy to target the correct IP/interface within the "backend configuration". Testing with Reef has shown that the IPs defined within the backend are the IPs that resolve to the host which in practice would potentially place the nfs traffic on a management network. In my lab the management network is 25Gb, but I think that it is more usual for the mgmt network to be 1Gb which would be severly limiting to the potential performance of Ganesha.
The problem appears to be related to the following;
Perhaps the IP address for Ganesha should be specified in a separate field in the spec, to ensure that the backend is routed correctly?
Updated by Bartosz Bezak about 1 month ago
the same applies to rgw based ingress, ip to HAProxy config is taken from management network:
https://github.com/ceph/ceph/blob/v18.2.2/src/pybind/mgr/cephadm/services/ingress.py#L145-L148
https://github.com/ceph/ceph/blob/v18.2.2/src/pybind/mgr/cephadm/templates/services/ingress/haproxy.cfg.j2#L76-L77
Updated by Bartosz Bezak about 1 month ago
Bartosz Bezak wrote:
the same applies to rgw based ingress, ip to HAProxy config is taken from management network:
https://github.com/ceph/ceph/blob/v18.2.2/src/pybind/mgr/cephadm/services/ingress.py#L145-L148
https://github.com/ceph/ceph/blob/v18.2.2/src/pybind/mgr/cephadm/templates/services/ingress/haproxy.cfg.j2#L76-L77
https://github.com/ceph/ceph/blob/v18.2.2/src/pybind/mgr/cephadm/services/ingress.py#L167