Thank you very much @AndrewWilliams3142 !

Please feel free to add a whatsnew @AndrewWilliams3142

Hey both, I've added a test to check that dask doesn't compute when calling either xr.corr() or xr.cov(), and also that the end result is still a dask array. Let me know if there's anything I've missed though! thanks for the help :)

@dcherian, regarding the apply_ufunc approach, I might leave that for now but as you said it can always be a future PR

Thanks for that @dcherian ! I didn't know you could use print debugging on chunked operations like this!

One thing actually: If I change da = da.where(missing_vals) to da = da.where(~missing_vals) then we get the results we'd expect. Do you think this fixes the problem?

``` def _get_valid_values(da, other): da1, da2 = xr.align(da, other, join="outer", copy=False)

# 2. Ignore the nans
missing_vals = np.logical_or(da1.isnull(), da2.isnull())

if missing_vals.any():
    da = da.where(~missing_vals)
    return da
else:
    return da

```

print(da_a.map_blocks(_get_valid_values, args=[da_b]).compute()) <xarray.DataArray (space: 3, time: 4)> array([[1. , 2. , 3. , 4. ], [1. , 0.1, 0.2, 0.3], [2. , 3.2, nan, 1.8]]) Coordinates: * time (time) datetime64[ns] 2000-01-01 2000-01-02 2000-01-03 2000-01-04 * space (space) object 'IA' 'IL' 'IN' *

Well that was confusing! if missing_vals.any(): will not be triggered if all the values in a block are valid. With .chunk({"time": 1}), some of the blocks are all valid.

I discovered this with print debugging and the single-threaded scheduler which loops over blocks in a for-loop ```

Define function to use in map_blocks

def _get_valid_values(da, other): da1, da2 = xr.align(da, other, join="inner", copy=False)

# 2. Ignore the nans
missing_vals = np.logical_or(da1.isnull(), da2.isnull())
print(missing_vals)

if missing_vals.any():
    da = da.where(missing_vals)
return da

da_a.map_blocks(_get_valid_values, args=[da_b]).compute(scheduler="sync") ```

For better performance, we should try a dask.array.map_blocks approach with duck_array_ops.isnull

4559 and #4668 fixed a similar issue using map_blocks, maybe you can use that as a reference?

Hi @dcherian , just thinking about your suggestion for using map_blocks on the actual valid_values check. I've tested this and was wondering if you could maybe point to where I'm going wrong? It does mask out some of the values in a lazy way, but not the correct ones.

```python3 da_a = xr.DataArray( np.array([[1, 2, 3, 4], [1, 0.1, 0.2, 0.3], [2, 3.2, 0.6, 1.8]]), dims=("space", "time"), coords=[ ("space", ["IA", "IL", "IN"]), ("time", pd.date_range("2000-01-01", freq="1D", periods=4)), ], ).chunk({'time':1})

da_b = xr.DataArray( np.array([[0.2, 0.4, 0.6, 2], [15, 10, 5, 1], [1, 3.2, np.nan, 1.8]]), dims=("space", "time"), coords=[ ("space", ["IA", "IL", "IN"]), ("time", pd.date_range("2000-01-01", freq="1D", periods=4)), ], ).chunk({'time':1})

print(da_a)

<xarray.DataArray (space: 3, time: 4)> array([[1. , 2. , 3. , 4. ], [1. , 0.1, 0.2, 0.3], [2. , 3.2, 0.6, 1.8]]) Coordinates: * space (space) <U2 'IA' 'IL' 'IN' * time (time) datetime64[ns] 2000-01-01 2000-01-02 2000-01-03 2000-01-04

print(da_b)

<xarray.DataArray (space: 3, time: 4)> array([[ 0.2, 0.4, 0.6, 2. ], [15. , 10. , 5. , 1. ], [ 1. , 3.2, nan, 1.8]]) Coordinates: * space (space) <U2 'IA' 'IL' 'IN' * time (time) datetime64[ns] 2000-01-01 2000-01-02 2000-01-03 2000-01-04

Define function to use in map_blocks

def _get_valid_values(da, other): da1, da2 = xr.align(da, other, join="inner", copy=False)

# 2. Ignore the nans
missing_vals = np.logical_or(da1.isnull(), da2.isnull())

if missing_vals.any():
    da = da.where(missing_vals)
    return da
else:
    return da

test

outp = da_a.map_blocks(_get_valid_values, args=[da_b])

print(outp.compute())

<xarray.DataArray (space: 3, time: 4)> array([[1. , 2. , nan, 4. ], [1. , 0.1, nan, 0.3], [2. , 3.2, 0.6, 1.8]]) Coordinates: * time (time) datetime64[ns] 2000-01-01 2000-01-02 2000-01-03 2000-01-04 * space (space) object 'IA' 'IL' 'IN' ```

Hi @AndrewWilliams3142 , thanks for another PR!

This looks good. Could we add a test like https://github.com/pydata/xarray/pull/4559/files#diff-74d2dc289aa601b2de094fb3a3b687fd65963401b51b95cc5e0afcd06cc4cb82R45? And maybe reference this or #4559 as an explanation of what's going on.