Ok so running the tests from a script in the tests directory

from test_flux_xr import Test_xarray

test_class = Test_xarray()

shape = (10, 13, 12)
chunks = {"dim_0": 1}
skin_correction = True
test_class.test_chunked(shape, chunks, skin_correction)

gives me these sort of errors:

 ===================================================================
                    ----- AeroBulk_init -----

     *** Bulk parameterization to be used => "coare3p0"
        ==> will use the Cool-skin & Warm-ayer scheme of `coare3p0` !
     *** Computational domain shape: Ni x Nj = 00001 x 00013
     *** Number of time records that will be treated:          12
     *** Number of iterations in bulk algos: nb_iter  =    6
     *** Filling the `mask` array...
         ==> no points need to be masked! :)
     *** Type of prescribed air humidity  `relative humidity [%]`
 ===================================================================
 *** E R R O R :
  COARE3P0_INIT => allocation of Tau_ac, Qnt_ac, dT_wl & Hz_wl failed!

switching the algo shows a similar thing for ecmwf:

 *** E R R O R :
  ECMWF_INIT => allocation of dT_wl & Hz_wl failed!

These errors go back to the aerobulk code here and here, both parts of the codebase seem to allocate some internal optional output variable.

Does anyone have a hint to why this would not sit well with threadsafe (cc @rabernat)?

This might be relevant: https://stackoverflow.com/questions/34579769/f2py-error-with-allocatable-arrays

Quick update. This only applies when using dask arrays, interestingly. The tests with the numpy only wrappers in tests/test_fortran.py all pass!

Is that because Dask uses threads?

Also, what is different about skin that makes it not thread safe vs. noskin.

Sorry the previous comment was circular:
As far as I can tell the difference is in the allocation of optional output which is only done when chosing the skin correction option (there are slight differences between the actual algos but the failure happens at the same ALLOCATE call in both, see below:

These errors go back to the aerobulk code here and here, both parts of the codebase seem to allocate some internal optional output variable.

Is that because Dask uses threads?

Any of my dask test cases has used the threaded scheduler (default). Before setting threadsafe, this worked for either skin/noskin but was very slow (basically running in serial). After setting threadsafe the noskin wrapper parallelizes as expected but the skin version actually FAILS.

Note that once #40 is merged, the skin code is no longer "threadsafe". The "threadsafe" option was removed in that PR to allow for tests to pass.

Ill close this for now, since we have decided to move on without threadsafe for the skin option for now. We can reopen if needed