Yes but could you break it up into chunks of sets of gradients to compute? I know that compute needs the full chunk to compute a set. Again, things I’m exploring but ultimately no different than just having the full dataset on disk and just scaling out compute nodes in ro mode.