STDIN prompts rendering is nondeterministic with respect to stream outputs

[CSEMike]

I believe the Jupyter FE is receiving STDIN and stream messages in a nondeterministic order.

Here's a small repro:

import getpass, sys
print (2)
sys.stdout.flush()
_ = getpass.getpass()

I would expect this to render 2 before the getpass prompt in all executions. Yet, if I execute this a few times, I occasionally see the reverse order. (Perhaps 1-2 instances in 10 executions.)

Example screenshot:

My conjecture is that what is happening is that STDIN channel messages and IOPUB messages are not deterministically ordered for a given cell execution.

[takluyver]

I think your conjecture is on the mark. I don't know if there's an easy way to solve this, though.

[Kai-Chen]

I'd like to take a look at what's causing this problem. Any hint at which part of the code base should I look into?

[jasongrout]

The first thing I'd do is try to reproduce the problem and look in the browser debugger to check the order of messages coming into the browser. If it is indeed the message order that is messed up, then the problem is likely in the python side. If the order of the messages is okay, but we still see them in the wrong order, the problem is in the javascript.

[Kai-Chen]

I was having trouble reproducing the problem, till I added a clear_output() call (see sample code below).

And it's the server messages that's out of order (see attached screenshots)

Which part of the python code should I look into next?

Sample code

from IPython.display import clear_output

def get_player_input(banner):
    while True:
        clear_output()
        print(banner)
        choice = input(f'(H)it, (S)tand, or (Q)uit: ').upper()
        if choice == 'H' or choice == 'S' or choice == 'Q':
            return choice

choice = ''
while choice != 'Q':
    choice = get_player_input('Welcome to Blackjack!')
    print(f'You chose {choice}')

Message Order

• input_request message arrives first (should be second)

• stream message arrives second (should be first)

System Information

kai@sccdev01(0)% uname -a
Linux sccdev01 4.18.12-arch1-1-ARCH #1 SMP PREEMPT Thu Oct 4 01:01:27 UTC 2018 x86_64 GNU/Linux
kai@sccdev01(0)% jupyter-notebook --version
5.5.0
kai@sccdev01(0)% python --version
Python 3.7.0

[sergenti]

any news? this is so weird

[jasongrout]

Interestingly, I can reproduce the issue with Michael's code on classic notebook (happened about 4 times in 100 tries), but I executed his code block 150 times on JupyterLab and didn't see the issue at all.

I think Michael correctly diagnosed the issue, and my guess is that the root problem is that messages on different zmq channels are not deterministically ordered between channels. One idea: since each message has a timestamp, perhaps we can use the timestamp to guide us in ordering messages across different zmq channels from the same kernel (if the timestamp has high enough resolution). Or more deterministically, we could introduce a message counter in message metadata.