@vectorize'd functions can't be pickled

numba 0.37.0, Python 3.6, Anaconda 5, Linux 64-bit

Being able to pickle (the reference to) a function is a fundamental prerequisite to dask.distributed. Functions wrapped with @vectorize fail to pickle, whereas @guvectorize has no issues. I’ve tried both with and without explicit signatures.

From the documentation for @vectorize:

If signatures is non-empty, then the decorator will compile the user Python function into a Numpy ufunc. If no signatures are given, then the decorator will wrap the user Python function in a DUFunc instance

However, it looks like the above is not true - @vectorize always returns a DUFunc, whereas @guvectorize returns a ufunc:

import pickle
from numba import vectorize, guvectorize

@vectorize(["float64(float64)"])
def inc1(x):
    return x + 1
    
@guvectorize(["void(float64, float64[:])"], "()->()")
def inc2(x, out):
    out[0] = x + 1
   
print(type(inc2))
print(pickle.dumps(inc2))
print(type(inc1))
print(pickle.dumps(inc1))

Output:

<class 'numpy.ufunc'>
b'\x80\x03cnumpy.core\n_ufunc_reconstruct\nq\x00X\x0b\x00\x00\x00__mp_main__q\x01X\x04\x00\x00\x00inc2q\x02\x86q\x03Rq\x04.'
<class 'numba.npyufunc.dufunc.DUFunc'>
Traceback (most recent call last):
  File "vectorize_pickle.py", line 15, in <module>
    print(pickle.dumps(inc1))
TypeError: can't pickle DUFunc objects

[stuartarchibald]

It will be fixed, but I am not sure when. We’re aware that jitclass needs some work, we added a tag a while back to group all the things may need a fix https://github.com/numba/numba/issues?q=is%3Aissue+is%3Aopen+label%3Ajitclass, there’s quite a few 😃

[dargen3]

Thank you for quick response. Is there any estimation, whether (and possibly when) it will be fixed? But i absolutely understand, if it has low preference.