I came up with the idea for this feature a few days ago and figured it would be worth a shot. Since the only way to subset datasets before loading them into memory is by downloading certain variables, I wanted some capability to preprocess the datasets before they're loaded, like the preprocess keyword argument in xarray.open_mfdataset(). I think this could be especially useful for loading in multiple files at once using FastHerbie -- this can become problematic if you've got global data for many forecast times. The most obvious usage would be spatial subsetting, but it could also be used for dropping variables and dimensions that could cause dataset merge conflicts. Please have a look and see if you think it accomplishes what I think it does!