apply_ufunc with dask="parallelized" and "allowed"

[dougiesquire]

A large number of xskillscore methods use xarray's apply_ufunc with dask="parallelized" for dask array support. A preferred option if the wrapped function natively supports dask arrays is to use dask="allowed". See here for details.

This issues list all current methods within xskillscore that use apply_ufunc and tries to summarise for each method how much work is involved in enabling dask="allowed".

xskillscore.contingency

• gerrity_score : already dask="allowed"

xskillscore.deterministic

• linslope : all numpy functions used can ingest and return dask arrays so can simply switch dask="parallelized" for dask="allowed"
• pearson_r : all numpy functions used can ingest and return dask arrays so can simply switch dask="parallelized" for dask="allowed"
• pearson_r_p_value : requires slight refactor to how masked nans are reset here since dask doesn't seem to be able to take the length of a empty array
• effective_sample_size : all numpy functions used can ingest and return dask arrays so can simply switch dask="parallelized" for dask="allowed"
• pearson_r_eff_p_value : requires slight refactor to how masked nans are reset here since dask doesn't seem to be able to take the length of a empty array
• spearman_r : need to wrap bottleneck.nanrankdata with dask.map_blocks or equivalent, although I don't know if this is any better than using dask="parallelized"
• spearman_r_p_value : as above
• spearman_r_eff_p_value : as above
• r2 : all numpy functions used can ingest and return dask arrays so can simply switch dask="parallelized" for dask="allowed"
• me : all numpy functions used can ingest and return dask arrays so can simply switch dask="parallelized" for dask="allowed"
• rmse : all numpy functions used can ingest and return dask arrays so can simply switch dask="parallelized" for dask="allowed"
• mse : all numpy functions used can ingest and return dask arrays so can simply switch dask="parallelized" for dask="allowed"
• mae : all numpy functions used can ingest and return dask arrays so can simply switch dask="parallelized" for dask="allowed"
• median_absolute_error : all numpy functions used can ingest and return dask arrays so can simply switch dask="parallelized" for dask="allowed"
• mape : all numpy functions used can ingest and return dask arrays so can simply switch dask="parallelized" for dask="allowed"
• smape : all numpy functions used can ingest and return dask arrays so can simply switch dask="parallelized" for dask="allowed"

xskillscore.probabilistic

• crps_gaussian : this is a wrapper on properscoring which expects numpy arrays (and actually triggers compute with dask="allowed"). Getting dask="allowed" working properly would require a full refactor of properscoring
• crps_quadrature : this is a wrapper on properscoring which expects numpy arrays (and actually triggers compute with dask="allowed"). Getting dask="allowed" working properly would require a full refactor of properscoring
• crps_ensemble : this is a wrapper on properscoring which expects numpy arrays (and actually triggers compute with dask="allowed"). Getting dask="allowed" working properly would require a full refactor of properscoring
• brier_score : this is a wrapper on properscoring which expects numpy arrays (and actually triggers compute with dask="allowed"). Getting dask="allowed" working properly would require a full refactor of properscoring
• threshold_brier_score : this is a wrapper on properscoring which expects numpy arrays (and actually triggers compute with dask="allowed"). Getting dask="allowed" working properly would require a full refactor of properscoring
• rank_histogram : need to wrap bottleneck.nanrankdata with dask.map_blocks or equivalent, although I don't know if this is any better than using dask="parallelized"
• reliability : already dask="allowed"

xskillscore.resampling

• resample_iterations_idx : use dask moveaxis when dask array. This would be easily handled with duck array ops - see below.

[dougiesquire]

As a general note, I'd suggest that we implement in xskillscore something like xarray's duck_array_ops module. This would make some of the suggestions above very easy to implement/read and replace, for example, the _get_numpy_funcs function in xskillscore.deterministic. I'll try to open a PR for this when I next find some time.

[dougiesquire]

I should also point out that even for those functions that ostensibly can just be switched to dask="allowed", I think @ahuang11 encountered some issues when the forecasts, observations and weights are not all dask or all numpy arrays. Details here. We'll need to resolve these issues with the first PR.