How can I prevent the jupyter kernel from buffering stdout/stderr?

Kernels
1

I am using an AsyncKernelManager + AsyncKernelClient to execute code in the jupyter kernel. I notice that even if code executes instantly, stdout/stderr messages both take much longer to appear on the iopub channel than the actual code executing.

Is the Jupyter kernel doing some sort of buffering internally?

2

The protocol doesn’t say the granularity of stream messages, it’s up to the kernel.
ipykernel at least defaults to buffering up to 200msec before flushing a stream message:

You can read/set sys.stdout.flush_interval and sys.stderr.flush_interval, even from a notebook, to tune this.
With a client, something like this probably works:

code='''
import sys
sys.stdout.flush_interval = 0.01
sys.stderr.flush_interval = 0.01
'''
client.execute_interactive(code=code)  # and optionally output_hook to suppress logging