As long as you treat servers as cattle, you can use Hetzner's own load balancer service and then you don't have a SPOF that you manage yourself. Their LBs are advertised as redundant / fault tolerant.
I don't know if it is possible to treat "Serverbörse" servers as cattle. They are all different. I know that k8s and Docker Swarm could in theory balance load between different machines, but never tried it in practice. But I had in practice some weird glitches with different CPU/motherboards/memory.
Also, comment I was replying to mentioned a 48 euro budget, it is a price of a single server.
yes, it may be bad for indiscriminate/dumb load balancing. then again this only works in the firstplace if the work units represent a somewhat equivalent and small-ish workload.
once you place value on determinism (in regards to time spent on a task) you want a tightly specced distribution mechanism and/or a feedback loop to communicate busystate back to the LB.
I have two of the AX41-NVMe, and only one was lucky enough to get a "Intel Corporation I210 Gigabit Network Connection (rev 03)" which has the hardware timestamping.
In the finland datacenter, it appears there is a PTP running, though the offset is ~2.33 seconds off from NTP. Chrony says its a false-ticker, but I haven't really figured out how to get it configured correctly nor have I asked Hetzner for help. I've mostly just played around with it.
I did manage to also accidentally discover a local ISP's stratum 1's server is extremely close to me, as in a few microseconds away (I accidentally put hetzner's servers as a pool instead of a server and my NTP 'discovered' the stratum 1) ... I'm not using it, but I've thought about reaching out to them to ask if I can use it if I'm very nice.
Also, comment I was replying to mentioned a 48 euro budget, it is a price of a single server.