Keep in memory original data type for writing

[remi-braun]

It could be very nice to keep in memory the original data type (ie. uint16 for landsat data or uint8 for classified data).

Indeed, for now (unless I missed something), if we open a dataset with masked=True and the data is casted to float, we lose the information of the original datatype.
So we always write the dataset on disk in float, even when it is not justified (when the nodata is replaced by its true value)

[snowman2]

That is a really good idea. I think this discussion should start with xarray to ensure consistency.

[snowman2]

Seems related:

• NetCDF encoded data not automatically decoded back into original dtype pydata/xarray#4973
• Masking and preserving int type pydata/xarray#3955

[snowman2]

This seems more relevant: pydata/xarray#1194

[dcherian]

You should be able to set .encoding["dtype"] at write time.

I bet it's being saved in encoding at read time but then is lost later.

[snowman2]

Fix in #307