Support for understanding xarray and cfgrib memory usage

[meteoDaniel]

Dear guys, I am having some advanced questions to understand how xarray works.

1. As I understand open_dataset and especially open_mfdataset correctly , xarray just reads the file headers? So that the whole data is not loaded to the memory?
2. In case you know the indices you want to access, is xarray able to return only the required data without loading the whole array into memory e.g. with.isel?
3. I assume loading the whole dataset into memory is the bottle neck or am I wrong?
4. Does this already apply to grib files?
5. Do you see any advantages of using grib data instead of netcdf to archive thousands of datasets ?

PS: This example mention such a behaviour using netcd4 library: https://annefou.github.io/metos_python/07-LargeFiles/