Handling of local time

[frizwi]

Xarray converts time to UTC when TZ is present in the source. This makes it's way to various plotting routines, eg. transect.

ds = emsarray.open_dataset('https://thredds.nci.org.au/thredds/dodsC/fx3/gbr4_H4p0_ABARRAr2_OBRAN2020_FG2Gv3_Dhnd/gbr4_simple_2022-10-01.nc')
ds['time'][0].data

array('2022-09-30T14:00:00.000000000', dtype='datetime64[ns]')

Users typically expect the timepoints in the netcdf file to be shown as is.

So this ticket is a request for an additional argument to emsarray.open_dataset() for make_time_local.

[mx-moth]

Unfortunately the situation isn't quite this simple.

emsarray.open_dataset() is not the right place to make these transformations as a dataset can come from many different sources, so sometimes users would be able to set make_time_local=True, and other times they would have to make the fix manually.

numpy datetimes do not have any timezone information associated with them. xarray assumes all times are in UTC. This matters when saving datasets. Opening a dataset, updating the datetimes to apply a timezone offset, and then saving the dataset again would result in the new local time values being interpreted by xarray as being UTC when encoding them. The dataset on disk would then have different datetime values than the original dataset, and opening the dataset again would load the local datetime values as if they were UTC. It all gets confused. There is no way of informing xarray that the datetime values in a DataArray are in a different timezone.

I agree that the behaviour of xarray is not ideal, and "modifying" the values as they appear in the dataset is confusing especially when the dates in question align nicely in the local time. Changing the default behaviour of xarray can add to this confusion though as now some datasets will be in UTC and some in localtime depending on how they were opened.

I can see the value in a method that shifts all UTC datetimes by the timezone offset stored in the units attribute. I have added such a method to the ncwps for this exact purpose. This method would have to have some large caveats that it is not round trip safe when saving the dataset again.

One possible alternative is adding a timezone parameter to the plot.make_title() function that will adjust the timezone of any datetimes appropriately. It could have values of:

• datetime.utc for the current behaviour, as the default,
• any datetime.tzinfo or datetime.timezone value to present the time in that timezone specifically, and
• literal "data" to present the datetimes in the timezone encoded in the units parameter.

This parameter would be passed through from the appropriate higher level plot functions.