The q came up about cloud based vcpus and threading.
These are problematic as far as I recall. One should make sure that the vcpus are always on a single physical machine. So maybe it’s best to do some benchmarks to find out how virtual cpus are mapped to actual physical ones.