I wish that they had run more benchmarks beyond the 20k idle connections one. While that case is important, they didn’t demonstrate how the system behaves in, for example, a 2000 person chat ro, a-la slack.