on large-ish dataset the current MDS is running out of memory. These compromises seem promising: https://arxiv.org/pdf/2007.11919.pdf