BUG: scipy.special.stdtrit is not thread-safe for df.size > 500 #18678
Description
I encountered a very weird issue, but I can reproduce it reliably. I encountered it first on a virtual Ubuntu machine and can reproduce it on my AArch64 Mac, so I am hopeful you will be able to reproduce it as well. I could not build scipy from source, but the issue is the same from versions 1.9.2 to 1.11.0rc on Python 3.11. It is reproducible as follows:
import numpy as np
import scipy.special
import threading
q = 0.95
df = np.arange(5, 506)
def calculation():
for i in range(100):
print(f"Calculation number {i}.")
x = scipy.special.stdtrit(df, q)
if __name__ == "__main__":
thread1 = threading.Thread(target=calculation)
thread2 = threading.Thread(target=calculation)
thread1.start()
thread2.start()
thread1.join()
thread2.join()This always hangs Python completely. Output is for example:
Calculation number 0.
Calculation number 1.
Calculation number 2.
Calculation number 3.
Calculation number 4.
Calculation number 0.
With df = np.arange(5, 505) the program always finishes successfully. My workaround now is to guard calculations involving stdtrit (relevant for me is scipy.stats.t.ppf) with a threading.Lock, which works. When replacing the threads with processes, the issue is also not present.