As a happy conclusion to the previous post, it seems that updating the version of Ubuntu to 24.04 resolved the outstanding issues.
When the hardware was initially commissioned 22.04 was the current LTS version, and I’m a little wary of using an LTS version which is only a month or two out of the gate, so I didn’t immediately switch.
As it became clear that were still some serious issues with the X710 cards which we were unable to fully mitigate then rebuilding the servers with Noble Numbat seemed like the only sensible choice.
Subsequently, I’ve not seen an issue with XDP hooks being lost when renegotiating links with the switches, and the servers are each happily handling hundreds of megabits per second of ingress traffic (which corresponds to tens of gigabits/s of data served to the clients - DSR is simply awesome) without a hitch, including LACP/MLAG. I guess there was something in the driver module which received attention in later kernels.
There was a brief squeaky bum time moment after bringing the new hardware into production when connections were failing at peak hours, but this turned out to a be a traffic policer policy which had not been updated to include a new virtual address for a service during migration.
I’m still a little suspicious of the X710 card given other’s bad experiences with it, and I would give serious consideration to auditioning a Broadcom chipset if we were deploying a greenfield datacentre, but as everything has been quiet so far over the Christmas/New Year change freeze period (which co-incides with a couple of weeks of annual leave), I’m sanguine enough.