Lazy loading of wide dataset

I’m not familiar with the internal implementation of the xarray, but what seems to be different is the shuffling. Could you test your simple code snippet with random indices xarray[0:64,:].values instead of contiguous ones and compare the loading speed?