@dcherian It works for the very simple test I did. We have fewer groupby().quantile() than you might think because the main "grouped quantile" function we implement is meant for "day-of-year" grouping including a window around each doy and the whole process is done without calling groupby. That code predates many improvements in xarray, I guess we should now revisit it, when this PR is completed!

One test in xclim's CI where we perform a grouped quantile failed when I modified it to use a chunked array : ValueError: Aggregation nanquantile is only supported for method='blockwise', but the chunking is not right. . I understand the issue and the code did run after a call to flox.xarray.rechunk_for_blockwise. I guess that's seen as a user's responsability here ? Has there be any discussions on automatically running rechunk_for_blockwise ?

It should do that if you explicitly pass method="blockwise".

But great catch! I'll have to add a fall back. I haven't thought about how to infer that rechunk_to_blockwise is sensible.

the main "grouped quantile" function we implement is meant for "day-of-year" grouping including a window around each doy and the whole process is done without calling groupby.

When you do this, is the time axis a single chunk?

When you do this, is the time axis a single chunk?

That's the "custom implementation" part : we manipulate the array and rechunk so that all members of one group (a day of year plus a window) fall into the same chunk.

Can you link to that section of the code please?

The function is here : https://github.com/Ouranosinc/xclim/blob/acbccb8902fb05318c53d76beb1c97067dfb8f11/xclim/core/calendar.py#L627
The grouping is done through a custom multiindex and stacking, and thus it only supports daily data (technically is supports coarser but the function really is meant for daily).

I'll merge tomorrow if there are no comments.