Re: [pve-devel] [PATCH frr 0/2] Bump FRR to 10.4.1

[Gabriel Goller <g.goller@proxmox.com>]

On 15.12.2025 08:57, Thomas Lamprecht wrote:
[snip]
> > FRR updates always had regression, when I had implement evpn I had
> > regression with 7.5,7.6,7.8,8.0,8.1  as far I remember.
> > 
> > so, you really don't need to big gap to have problem, it's occur all
> > time (I don't known, they don't seem to have automatic test on readl
> > infra)
> 
> Maybe something Gabriel could look into, as I'd prefer upstream testing
> more over us testing more, while both are required, the former helps every
> FRR user and should avoid some redundant effort.

Testing frr is inherently difficult. Upstream has extensive topotests 
(integration tests with containerized network infrastructures) that are 
mandatory for all new features and bugfixes. However, routing protocols are 
(at least in frr) unfortunately non-deterministic -- even bgp route
selection can vary between runs.

For example, these are the recent frr regressions we've encountered
(since 8.5) (that I know about -- could be that I forgot some):

 - BGP BFD peers failing to establish sessions (bfd bug, backported by
   Stefan)
 - EVPN type-5 routes not being installed (bgpd race condition,
   backported by me)

Both have not been catched by the upstream integration tests and both
where not easily reproducable (flaky). They don't even have a topotest
right now, because we weren't able to write one that makes sense
(without inserting random sleeps or lua scripts that sleep inside of
frr).

Nevertheless we're trying to write some integration tests using full PVE
clusters with specific SDN topologies. While these probably won't catch
frr bugs directly, they provide us a reference to test against.

> > here we go again....  :/
> > 
> > https://forum.proxmox.com/threads/frr-update-to-10-4-1-1-broke-external-routing.177736/
> 
> This specific issue seems to have been our fault though :/

Looks like it was the users fault :)

> >>> The maintainers have told me this should improve going forward.
> > 
> > don't trust them ;)
> 
> Hmm, I really get that you have been burned a bit too often by them,
> but IMO the only good possibility to improve on their releases for our
> users in the mid/longterm is to more frequently update in smaller steps,
> and then give feedback or (cherry-pick) fixes back. And with Gabriel we
> now have a depth that is a bit more involved with upstream, so I have more
> confidence in that working out, albeit I certainly do not expect wonders
> (quickly).
> 
> In any case, we'll certainly move FRR updates, especially those for new
> releases along more slowly.


_______________________________________________
pve-devel mailing list
pve-devel@lists.proxmox.com
https://lists.proxmox.com/cgi-bin/mailman/listinfo/pve-devel