Tracing generator functions

[danem]

Weave Version: 0.51.28
Python version: 3.12.8

I've found that functions returning a generator aren't properly traced.

For example:

@weave.op()
def b():
  return 1


@weave.op()
def a():
  yield b() + 1

weave.init("test")
for x in a():
    print(x)

Results in two separate traces:

While this works properly:

@weave.op()
def b():
  return 1


@weave.op()
def a():
    return b() + 1

This is relevant for tracing functions that stream results back to the user.

[jamie-rasmussen]

Thanks for reporting, we have a corresponding issue in our internal tracker, WB-19801.

[danem]

Thanks! Any idea when it might be addressed? I'm evaluating different observability tools, and while this isn't a deal breaker or anything it is rather annoying given our usage.

[jamie-rasmussen]

I can tell you that this hasn't been selected for development in the next few weeks, but we can't yet commit to a time frame for when it might be. I can also say that a number of users have mentioned it, which does help with its prioritization.

[jdai1]

@danem Hey! I'm also in the same boat as you here — the streaming thing is kind of annoying. I'm wondering if you had any super positive experiences with other LLM observability tools relative to Weave?

[nataliachodelski]

@jamie-rasmussen - I found this recent PR from Data Dog's python client in which they addressed issues they had decorating sync and async generator functions and thought it might be helpful to the weave team re: fixing this bug.

DataDog/dd-trace-py#11413

it looks like the function they're using to determine if a function returns a generator is isgeneratorfunction, which is coming from from inspect import isgeneratorfunction via this file: https://github.com/DataDog/dd-trace-py/blob/main/ddtrace/internal/compat.py