Ordinarily numba jitted functions cannot dynamically access global variables - the value is fixed as of compile time. A pythonic behaviour can be forced by forcing objmode. (There is a reduction in performance for using objmode).

from numba import njit, objmode 
@njit 
def access_global_variable_1():
    return global_variable

@njit
def access_global_variable_2():
    with objmode(retval='int64'):
        retval = globals()['global_variable']
    return retval

global_variable = 0
access_global_variable_1() # returns 0 
access_global_variable_2() # returns 0 

global_variable = 1
access_global_variable_1() # returns 0 
access_global_variable_2() # returns 1

However, this breaks when we parallelise functions

from numba import njit, objmode, prange
import numpy as np

@njit
def _add(a):
    with objmode():
        globals()['acc'] += a

@njit(parallel=True)
def parallel_sum_global(arr):
    for i in prange(len(arr)):
        _add(arr[i])

@njit(parallel=False)
def sum_global(arr):
    for i in prange(len(arr)):
        _add(arr[i])

@njit(parallel=True)
def parallel_sum_local(arr):
    acc = 0
    for i in prange(len(arr)):
        acc += arr[i]
    return acc

n = 100
print('True answer:', np.arange(n).sum()) # True answer: 4950 

acc = 0
parallel_sum_global(np.arange(n)) 
print('Numba parallel global answer:', acc) # Numba parallel answer: 78

acc = 0
sum_global(np.arange(n)) 
print('Numba global answer:', acc) # Numba global answer: 4950

acc = parallel_sum_local(np.arange(n))
print('Numba parallel local answer:', acc) # Numba parallel local answer: 4950

It can't be a problem relating to accessing global variables since sum_global works as intended. It can't be a problem relating to parallelisation as sum_parallel_local works as intended. It is only a problem that occurs when both accessing global variables and parallelising.

Investigating more closely, parallel_sum_global correctly handles the first 12 array items but not after 12. I am developing on an 8 core computer.

Has anyone encountered this issue? Are there any workarounds?

(Disclaimer: I came across this issue working on something much more complicated, this is a minimal reproducer on a contrived example of summing arrays)