Hi Maria! All...
Are you looking for a failover or for load balancing?
I believe the best definition is "sharding," but with failover for shared resources.
BGP clients that, due to geographic distribution, naturally operate under a route reflector, but which can, if necessary, fail over to another unit in the event of some type of failure.
The problem arises when an event causes a specific set of downstream
clients to switch from one Route Reflector (RR) node to another. […] The
conclusion I’ve reached is that the healthiest approach would be for a
BGP session to be resumed from a different node, in such a way that the
BGP client remains unaware that this has occurred.
I’m kinda afraid that resuming long-lived TCP sessions from a
different node is quite a crazy thing to try to achieve.
Agreed!
[…] I am aware that “graceful restart” is a prerequisite for this
kind of implementation.
I would say it’s orthogonal. Graceful restart is useful for
situations where the BGP session ends.
Agreed!
And I’ve been reflecting... My hunch is that I don’t actually need stateful BGP session failover.
If I have a reliable mechanism to determine which node is the "active" one for a given resource-consumer-group, I can simply configure that local node to treat the BGP peer as "active."
I imagine Corosync/Pacemaker could help me prevent a "split-brain" scenario for the BGP session.
Furthermore, if I mandate the use of LLGR (Long-Lived Graceful Restart), then when a BGP session shifts from one box to another, the GR mechanism will perform exactly as designed. This should allow me to avoid the "noise"—all those routes reconverging and being recalculated, sessions constantly going up and down, and the general turbulence that currently occurs.