(at least when running in docker, maybe that's the bottleneck, but I hesitated to run this on my machine directly)