Interesting and tough problems indeed,
@i386. I think confidence in the CPU+RAM+mainboard can be raised significantly with 96h of Prime95 torture testing. Not sure if Facebook is doing basically the same on an ongoing basis from the article you posted? ECC-RAM seems not optional. DDR5 with its "ECC-lite" built-in will have to be seen. No overclocking. Quality PSU with 5+yrs warranty, showing reasonably trusted component selection by OEM.
And then you will find out that Intel introduces regressions in microcode updates for security flaws of their silicon. Been bitten more than once. So never upgrade that, or at least not without running the MC update on a second identical system for a bunch of months first. Also buy late-stepping CPUs and not the B0 release version, and never the first batch from a new process.
With the unmitigated disaster that is the Intel I219 ethernet solution, and more recently bugs shown in 2.5G chips from them, I lost complete faith in Intel networking chips. If it just has work, I turn to Mellanox or Chelsio.
And before I forget, always keep your 3-2-1 backups current and test restore occasionally.