Metallb: BFD support

[creativie]

Is this a bug report or a feature request?:

Feature request

What happened:

Bidirectional Forwarding Detection not supported

What you expected to happen:

Detect faults between two forwarding engines connected by a link

Anything else we need to know?:

https://en.wikipedia.org/wiki/Bidirectional_Forwarding_Detection

Environment:

• MetalLB version: 0.7.3
• Kubernetes version: 1.13.2

[aojea]

I can share my experience with BFD implemented in software and ... simply don't use it, BFD operates in the order of miliseconds , there is a nice explanation here https://networkengineering.stackexchange.com/a/45926

I'm curious about your needs, are the routing protocol timers not enough?

[creativie]

I also have experience with software BGP
I use BGP with BFD in load balancers (between Nexus and linux servers), and with thousands of requests per second, fast traffic switching is necessary
Therefore, BGP timers are not enough

[dshcherb]

BGP timers is a control plane mechanism while BFD is used for data plane. In a _generic_ BGP deployment, control plane may be separated from the data plane and so failure to reach the next hop of a route is not the same as a BGP speaker failure.

More importantly, hold time is specified in seconds and cannot be less than that while for certain types of applications <200ms failover time is required to avoid noticeable impact. On hold timer expiry a BGP connection is closed and routes that have been advertised by a speaker associated with that connection are removed from a router's RIB & FIB, however, between the actual failure time and eviction time, ECMP load-sharing continues to work as before and some L4 traffic flows will be black-holed as hashing is static per L4 flow for a given routing table size and the failed route(s) will still be there and will be used for forwarding before eviction.

I can also see that the default hold time in the example config is 120 seconds which is a lot in terms of failure detection.

[danderson]

+1 for BFD, it would be nice. I'm not prioritizing it until I have proper e2e tests in place to make it safe to extend MetalLB, however. But, consider the request recorded and prioritized :)

[salanki]

A Linux kernel BFD implementation would be the best.

[Thoro]

Not sure if it's relevant for anyone here, but I did a BFD implementation in Go. Since it provides a grpc api, I'm sure someone could find a use for it.

https://github.com/Thoro/bfd