[SSI-devel] HA-LVS failover

[Brian J. W. <Bri...@hp...>]

Aneesh-

IIUC, HA-LVS does failover by maintaining TCP connection state on both
the active LVS director node and the backup. This allows transparent
failover of the LVS director without losing any state. Unfortunately, I
realized a potential problem with this approach: if you restart the
HA-LVS service (`/etc/init.d/ha-lvs restart`), you'll lose the state on
both nodes.

A solution to this problem would be to do failover like the centralized
OpenSSI kernel services. They do not maintain state on a backup node.
Instead, the backup node rebuilds state during failover by asking every
node in the cluster what partial state they have. For example, the
HA-CFS token manager fails over by asking every node what tokens they have.

If you did HA-LVS failover like this, then you could leverage it to
rebuild the state when restarting the service. I'm not sure how
important it is to make this work correctly right now, but I thought I'd
offer it as food for thought.

Brian