Sticky Route Selection (when choosing between two identical looking routes)?

[coofercat]

I have a setup where I need to provide two linux servers as "internet NAT gateways" to a private network. I can't do any sort of load balancing or IP sharing between them, so have used BGP on all the clients to determine which gateway to use (FRR to the rescue!).

Both gateways are exactly identical, except that they have different Internet IP addresses. Thus, if a client starts talking through one but switches routes mid-conversation, then the communication will fail because of the IP address change. In practice this doesn't seem to happen very often, but it presumably will, and will be tricky to diagnose. As far as the routing table is concerned, it looks like this:

default proto bgp metric 20
	nexthop via 10.0.0.10 dev ens3 weight 1
	nexthop via 10.0.0.11 dev ens3 weight 1

BGP sets both routes to be identical, and the OS picks one for the communications it provides. I'm not sure how/why, but presumably it's free to switch routes at any time.

What is my best strategy for avoiding this problem?

Things I've thought of:

• Set one server as the primary and other as the secondary. All traffic from all clients will only use the primary unless it fails, which isn't a terrible outcome, but it would be nicer to load balance a bit between them. Also, if the primary fails, everyone will failover to the secondary, but when the primary recovers, they'll all fail back (which isn't really necessary)
• Implement some sort of "stickiness" or hysteresis on route selection (which is sort of happening already, but more by luck than by design). That is, let BGP do whatever it wants (and so hopefully roughly load balance), but somehow set the thresholds for changing routes to be really high, so unless the chosen route has completely failed it won't switch to the other one.

I'm not sure how best to go about either of these strategies (if indeed the second is possible at all?). Any pointers would be much appreciated. It it helps at all, I've included the "client" BGP config below (which is pretty simple, it more or less just uses anything, except if it's about the network it's directly connected to):

frr version 7.2.1
frr defaults traditional
hostname myserver
no ipv6 forwarding
service integrated-vtysh-config
!
router bgp 65000
 neighbor 10.0.0.10 remote-as 65000
 neighbor 10.0.0.10 soft-reconfiguration inbound
 neighbor 10.0.0.10 distribute-list 1 in
 neighbor 10.0.0.11 remote-as 65000
 neighbor 10.0.0.11 soft-reconfiguration inbound
 neighbor 10.0.0.11 distribute-list 1 in
!
access-list 1 deny 10.0.0.0/24
access-list 1 permit any
!
line vty
!

[ton31337]

Usually, people set weight or local-preference to prefer a specific peer for outgoing traffic.