create_ensemble_on with calendar noleap

[juliettelavoie]

Setup Information

• xscen version: '0.7.20-beta'
• Python version: 3.10.6
• Operating System: linux

Description

I am trying to use create_ensemble_on in to_dataset.
I want to convert calendars to noleap. I get an error to use cftime, but that arg is not available through to_dataset. It is fixed to calendar == "default". Why is cftime only acceptable with default ? Is there a problem with adding a use_cftime_for_calendar arg to to_dataset ?

xscen/xscen/catalog.py

Line 493 in db7c976

calendar, use_cftime=(calendar == "default"), align_on="date"

Steps To Reproduce

cat = xs.ProjectCatalog('simulation.json')
cat.search(**dict(
        variable='tasmax', 
        source=['MIROC6', 'CanESM5'],
        processing_level='biasadjusted',
        bias_adjust_project=['ESPO-G6-R2'],
        experiment=['ssp245'],
        member='r1i1p1f1')).to_dataset(create_ensemble_on=["institution", "source"], calendar='noleap')

gives:

ValueError: Calendar 'noleap' is only valid with cftime. Try using `use_cftime=True`.

Additional context

Note that technically these two sources are already noleap in cftime. When I use to_dataset without converting to a calendar, the time axis becomes a numpy datetime (but february 29th are still missing.) This seems like a weird behavior to me.
It leads to issue, because xr.infer_freqdoesn't work anymore (outputs None). This means that default xclim checks on indicators fail.

Contribution

• I would be willing/able to open a Pull Request to address this bug.

[aulemahal]

Hum. haha what was I thinking ? Maybe its a typo and I actually meant calendar != 'default' ?

Because "default" is the name xclim (and only xclim) gives to the numpy datetime calendar.

On the other hand, if all datasets are using noleap and you are not getting a noleap calendar output, than that's an issue of catalog.to_dask think. Strange!

[juliettelavoie]

calendar != 'default' fixes my problem, I'll make the change in my partition PR.