The current implementation of xarray loads the whole dataset in memory, which can be very large depending on the gridding, area and time steps.
a lazy-loading implementation (open_dataset) or an implementation modifying the file in place (netCDF4) would be preferable.